508 skills, found in plain words

Verify a procedural Blender asset or a rig with parameter, connectivity and motion checks - list each exposed parameter with its range and test minimum, default and maximum for broken geometry, check road or path connectivity at junctions and the separation of vehicle and pedestrian paths, check rig weights, foot contact and mesh collision across the walk cycle, and report engine import and asset update behavior - each fix tied to a retest. Use when a creator builds a geometry-node road kit, procedural set or character rig and reports glitches at some settings or frames. Not for generating meshes, scene layout against a reference photo, or animation authoring.

Verified onClaude Code

Plan and verify a Blender scene with measurable checks - write down dimensions, units, camera path and the delivery budget first, compare a render to the reference image for composition, shape, material and lighting with the view used, check modular parts for matching dimensions, connections, collisions in every state and collection structure, and for web delivery compare file size, texture count, load time and frame rate to the target. Use when a creator builds a scene from a reference photo, a camera previz, a modular kit with door or object states, or a web-ready asset, and wants to know what passes. Not for generating 3D meshes or textures, rendering, or sculpting.

Verified onClaude Code

Check an AI-generated mesh, a retextured asset or a generated VFX element against numbers before it enters a scene or composite - triangle count against the budget, non-manifold edges, inverted normals, UV overlap, retexture silhouette and per-frame flicker, and plate match for color space, light direction, focus and motion - with a pass or fail row per check and frame ids. Use when a generated asset arrives with script stats or a sequence and the creator asks whether it can go in. Not for generating or retopologizing the asset, rendering, or the final acceptance decision, which is the artist's.

Verified onClaude Code

Correct spoken-audio noise, loudness and mixing with measured export verification. Use when the user requests cleanup, normalization or voice/music balance for an identified audio track or supplied measurements.

Produce Korean speech with a text-to-speech engine end to end - script prepared for reading, pronunciation fixes through the engine's dictionary or markup, synthesis in sentence-sized chunks, stitching with natural pauses, loudness to target, listening QA, and a record of the voice's licence. Use whenever Korean text is going to be read aloud by a TTS or AI voice - store or in-app announcements, narration, guides, ads, dubbing - including a quick question about whether a sentence with times, dates, floors, prices or months can go into TTS as written. Not for cloning a real person's voice.

Prepare long text for synthetic narration and proof the result - separate main text from footnotes, citations and bibliography as the user chooses and flag every dropped or leaked part, mark pronunciation-risk tokens (names, numbers, dates, abbreviations) before synthesis with the chosen readings, and after synthesis compare the source text with the transcript to list omitted, repeated and misread spans with positions, plus chapter order, split-file sequence and speed or emotion mismatches. Use when someone turns a book, article or long PDF into a TTS audiobook or listening file and gives the text, the choice of what to read, or a transcript of the audio. Not for Korean spoken-number rules or sentence chunking (korean-tts-production) or for choosing a voice.

Verified onClaude Code

Reconcile narration take validity and regeneration blockers in a scene ledger. Use when the user requests the impact of script, display or approved voice/settings changes on existing scene takes, alignment and dependent captions.

Verified onCodex

Design a coherent set of timed interface or video sound effects with source rights and event playback checks. Use when the user requests action feedback, notification sounds or event-timed effects and stingers.

Spar on a brand's sound before anything is recorded or generated - what the brand should sound like and where people hear it, two or three distinct sonic directions (instruments, tempo, voice, a short logo sound) with the case against each, and one cheap test. Use when the user asks what their brand, app or video series should sound like, or wants a sound logo, notification sound or music direction. Not for mixing or producing audio.

Verified onClaude Code

Write or convert a script for listening rather than reading - numbers, dates, units and English spelled the way they are spoken, sentences that fit one breath, marked pauses, and a pronunciation list for names - for narration, ads, guides or text-to-speech. Use when the user asks for a voice-over script, narration, a podcast intro or text meant to be read aloud by a person or a TTS engine. Not for subtitles timing or audio editing.

Check a rendered asset - a thumbnail, banner, carousel slide or deck page - from its screenshot for clipped or overlapping text, low contrast, off-brand colours or fonts, forbidden patterns and font licences, and report each defect with where it is. Use when the user asks to QA, review or approve a rendered image or slide before it is published. Not for critiquing the concept or rewriting the copy.

Turn a brand's colours, type, logo rules and forbidden patterns into one tokens.json that every asset reads, with contrast checked and every value traced to a source. Use when the user wants to fix a brand's design rules in one file, set up design tokens for assets or slides, or keeps getting off-brand outputs, or has a list of HEX values to give role-based names with a collision and contrast check. Not for building a full UI component library.

Turn a vague client request for design work into a list of its ambiguities, at most four prioritised questions, a default for each that the client can simply confirm, and a short reply email, before any design starts. Use when a client message is unclear about the deliverable, size, copy, audience, deadline or brand, and the designer needs a production checklist. Not for choosing a visual direction (use visual-concept-sparring) or for reviewing a finished asset.

Verified onClaude Code

Compare approved font repertoire and shaping controls before and after subsetting using supplied binary identities and named-shaper observations. Use when character coverage or a small valid font hides lost layout behavior. Not for designing fonts, brand typography selection or a new licensing audit.

Verified onCodex

Recover the text, hierarchy and colours of an AI-generated mockup or ad image into a reviewable text block, marking every uncertain or garbled character, listing fonts only as candidates and never correcting a word silently. Use when someone has a client mockup or ad image (or a description of one) and needs the copy retyped into an editable layout. Not for generating or redrawing the image, for rebuilding the layout file, or for QA of a finished asset (use asset-qa-review).

Verified onClaude Code

Spar on the visual direction of a campaign, page or asset before anything is drawn - ask the few questions that change the look, offer two or three distinct directions with the case against each, and narrow to one to test. Use when the user asks for a mood, a look and feel, a visual concept or design references for something not yet designed. Not for critiquing a finished design or generating images.

Verified onClaude Code

Check copy against one client's, brand's or season's rulebook - banned expressions, required phrasing, tone and format rules - so every finding points to the rule id it breaks, and keep each client's rulebook isolated so rules never leak from one client to another. Use when the user asks to review, proofread or approve copy, captions, banners or scripts for a specific client or brand guideline, or when an agency handles several clients with different rules. Not for general grammar editing or for ad-law claim review without a client rulebook.

Verified onCodex

Draft a proposal, report or press release from one client's own template, rules and decisions log - keep every earlier decision the client made, fill only facts the user supplied, and list every placeholder left unfilled instead of inventing numbers, quotes or dates - then check the draft against the template's required sections. Also builds a new proposal from earlier approved documents with a source line per section, stale facts flagged and old client names replaced. Use when the user asks to write a proposal, monthly report, press release or client document in a client's format, to reuse a client template for a new project, or to assemble a proposal from past approved proposals. Not for checking finished copy against a rulebook or for writing without a template.

Verified onClaude Code

Draft a person's own cover letter, resume edit, appeal, application essay, resignation or notice so it holds only the facts, dates and documents they supplied - list the supplied facts first, mark every missing fact as a placeholder, refuse to assert an unsupported circumstance, keep their voice, show a facts-used table, and for benefit, immigration or legal letters map each point to an evidence item and say a qualified person must verify. Use when an individual asks for a personal letter, appeal, essay or notice written for themselves. Not for commissioned ghostwriting for a client (ghostwriting-editorial), templated client documents (client-template-drafting) or a legal opinion.

Verified onClaude Code

Find and repair stock frames that AI drafts repeat in any language - the not-X-but-Y contrast, promotional openers and adjectives, verdict-style backchannels, choppy fragment runs, tangled modifiers - by counting the frame per paragraph first, keeping contrasts that carry a real claim, rewriting with the target language's own connectives and word order, adding no new facts, and re-counting afterwards. Use when a draft or chat reply keeps using one phrasing habit that an instruction to stop did not remove, or the user asks to make AI-sounding text read natural in German, Chinese, Japanese, French or English. Not for matching a client's voice, checking a persona, or rewriting Korean UI copy.

Verified onClaude Code

Edit or ghostwrite long-form text in a client's own voice - capture the voice from their samples, keep their claims and facts as theirs, flag anything that needs a source, track every change, and never present invented experiences as the author's. Also checks a draft against numbered voice rules taken from two samples and flags each drift. Use when writing a book chapter, essay, column or long blog post on someone's behalf, editing their draft, or checking whether a draft matches a client's voice. Not for short marketing copy or UI text, and not for scoring a persona's tone on test prompts (persona-tone-eval).

Verified onClaude Code

Write a spoken, read-aloud or fixed-size text (eulogy, toast, bedtime story, bio, short fiction, parody) to a time or character limit and a required-element list - convert the time to a word budget at about 130 words per minute, list the required elements and check each appears, count words or characters by script rather than by feel, trim from the end of each paragraph while keeping the elements, and add no invented facts about real people. Use when a text must run a set time aloud, fit a character or word cap, or include named elements. Not for scoring tone (persona-tone-eval) or editing a long manuscript (ghostwriting-editorial).

Verified onClaude Code

Plan and check a Naver blog post for search - the title, the opening, the structure, images and tags - without the keyword stuffing and copied text that Naver's ranking demotes. Use when the user writes, outlines or reviews a post for Naver Blog and wants it found in Naver search. Not for Google SEO, Tistory or web page meta tags.

Verified onClaude Code

Evaluate whether generated text matches a defined persona or brand voice - turn the voice into observable markers, build a small reference set, score outputs against a rubric with examples, and separate tone problems from factual or safety problems. Use when checking a chatbot persona, a brand voice prompt or a style-transfer model, or comparing two prompts for tone. Not for rewriting copy or general writing quality.

Verified onClaude Code

Turn a technical problem, outage or design issue into a note a non-technical executive, owner or homeowner can act on - impact, cause, action and next date stated separately in plain words, numbers, responsibility and assumptions kept exactly as the source gives them, overstated certainty and minimized risk flagged, and one decision the reader must make. Use when someone must explain an incident, outage, design issue or technical constraint to a reader who is not technical. Not for public status page updates (operations-cs status-incident-comms), the internal postmortem, or marketing copy.

Verified onClaude Code

Write release notes or support articles from developer notes, tickets, pull requests and diffs so that every claim cites a ticket, PR or note, unsupported claims are cut, UI and screen names appear only if they come from a screenshot, a string in code or the docs, and missing facts become questions instead of inventions. Use when someone asks for release notes, a changelog entry, a what's-new post or a support article from scattered dev notes, ticket titles or a diff. Not for marketing copy, ad claims or release notes for changes with no source material at all.

Verified onClaude Code

Write or fix Korean product copy - buttons, titles, body text, errors, empty states, notifications - to the public Toss writing principles, and say which principle each change follows. Use when the user asks to write, review or correct Korean UI text, microcopy or error messages. Not for blog posts, documents or English copy.

Verified onClaude Code

Review a Naver search ad account from its balance, actual spend and keyword spend, then compare bid alternatives as estimates kept apart from actual results. Use when the user asks where the ad money is leaking, how many days the budget will last, or what bid to set on Naver search ads. Not for changing bids, campaigns or budgets, and not for promising estimated clicks as results.

Verified onCodex

Check whether a pasted Korean blog or social post needs a sponsorship or ad disclosure, and whether the draft actually carries one, against bundled excerpts of the Korean fair labeling act and the endorsement guidelines. Use when the user asks whether a sponsored, gifted or affiliate post needs an ad label, or whether their disclosure wording and placement are enough. Not for a legal verdict, for drafting terms or privacy policies, or for real estate, finance or environmental ads.

Verified onCodex

Diagnose changes in the user's own Naver blog - visits, views, inflow sources, readers, dwell time, revisits and revenue - from the real statistics the datalab.tools extension reads. Use when the user asks why blog traffic or inflow dropped, wants last month's blog performance analysed, or asks which post is the problem. Not for forecasting future traffic, for blogs the user does not own, or for choosing the next topic.

Verified onCodex

Build Naver blog sidebar widgets, profile and introduction sections, recent-post lists and category menus as HTML that Naver accepts (span, br, hr, img and a only, inline styles, 170px and 2,000 bytes for widgets). Use when the user asks for a Naver blog widget, a sidebar profile card, a visitor counter or category menu HTML. Not for working widgets with scripts, hover effects, or installing anything on the blog.

Verified onCodex

Research Naver Cafe communities for a topic - how much cafe content exists for a keyword, which cafes talk about it most, and a given cafe's profile, boards and popular posts. Use when the user asks which cafe to join for a topic, whether cafe content on a keyword is saturated, what boards a cafe has, or what is popular in a cafe. Not for searching individual cafe posts (use the search-surface skill), not for member data, and not for posting or joining.

Plan a post or topic as a multi-page canvas - card news, an Instagram carousel or presentation slides - and build it page by page in the datalab.tools photo editor from gallery images and plain backgrounds, then check the render and hand the files over when asked. Use when the user asks for card news, a carousel, or slides made from a post or topic in the datalab.tools editor. Not for prompts to paste into outside image generators, for predicting saves or reach, or for filling slides with numbers the user did not give.

Verified onCodex

Check how Naver News comments moved - comment volume, number of commenters, hour of day, gender and age, device, country and news section spread - from the comment statistics the datalab.tools extension reads. Use when the user asks whether comment reaction grew, who comments and when, or which news sections a reaction spread to. Not for public opinion, sentiment or stance, which comment counts cannot show.

Verified onCodex

Check a Naver Smart Store in one pass - settlement, sales, order changes and product inspection issues - keeping each figure in its own aggregate. Use when the user asks how the store is doing today, why settlement differs from sales, or which store operations issues were missed. Not for processing, cancelling or shipping orders or editing products.

Verified onCodex

Benchmark another public Naver blog or the top blog posts for a keyword - profile, recent posts, popular posts, categories, visitor trend, top-post freshness and title patterns, and content gaps - from what the datalab.tools extension reads. Use when the user names a blog id to study, asks how a competitor blog is doing, or asks what the top posts for a keyword look like. Not for choosing the next topic, for the user's own blog statistics, or for predicting rank.

Rewrite the call to action that closes a blog post into two or three clear, specific candidates, each with the technique and the reason, using real search phrasing when the post promotes something. Use when the user asks to polish a post's closing line, fix a CTA, or write a purchase or sign-up prompt for a Naver blog post. Not for conversion predictions, A/B test design, invented social proof or urgency, or landing pages.

Verified onCodex

Reconcile creator group-buy contribution and settlement from supplied commission or resale terms. Use when checking a campaign's margin, refund exposure or net payout. Not for store-wide health, assumed tax, sales forecasts or executing orders and payments.

Verified onCodex

Research Naver influencer channels - find top influencers in a category, read a channel's profile, recent posts, challenges and public stats, and see which figures exist only for the owner's own channel. Use when the user asks who is doing well in a category, to analyse a competitor influencer, which of my influencer topics work, or what keywords are popular in an influencer category. Not for scoring or ranking influencers beyond what the tools return, estimating a channel's views or income, or finding blog data without the linked blogId.

Suggest the next blog topics for a Naver blog, each backed by search volume, trend and competing posts the datalab.tools extension checked, with varied angles and series ideas but no invented scores. Use when the user asks for topic ideas, the next thing to write, a series plan, or whether a topic can compete. Not for writing the post itself, for scheduling, or for definitive investment, medical or legal advice topics.

Find and call the right datalab.tools extension tool when the user does not know its name - for real Naver data (own blog, Place, shopping, search ads, Smart Store, Naver search, pumasi) and for controlling the datalab.tools post, photo and video editors - through the find, list, call and confirm-status tools. Use when a request needs live Naver data or an editor action and no more specific skill of this plugin covers it. Not for general knowledge, other platforms, or requests answered from pasted material alone.

Verified onCodex

Explain an actual neighbour's blog post and draft a personal reply grounded in the user's real questions and opinion. Use when the user struggles to understand a neighbour's post before replying. Not for relationship checks, invented agreement or experience, browser comment entry, submission or bulk replies.

Verified onCodex

Research other Naver Place listings and the local market - find a listing, nearby alternatives, notices, photos, coupons, smart store products, live broadcasts, pet-friendly places, the AI briefing, and pre-check a store name and representative keywords. Use when the user asks who the competitors nearby are, what a neighbouring store offers, whether a representative keyword will be blocked, or what is around a place. Not for the user's own reviews or reply queue (use datalab-place-reputation), and not for rankings, scores or sales estimates.

Check the user's own Naver Place listing - reviews, review sentiment, replies waiting, bookings, live waiting and coupons - and separate confirmed facts, priorities and reply drafts. Use when the user asks for a review analysis of their store, a reputation check, or which reviews to answer first. Not for inventing an overall reputation score, posting replies, or guessing customers' motives.

Verified onCodex

Change the post open in the datalab.tools Naver writing window with the native editor tools - text styles, headings, quotes, dividers, lists, tables, images, captions, link cards, attachments - and prove each change by reading the document back, then save it as a draft, list drafts, start a blank post or reopen a saved one. Use when the user says things like "turn this paragraph into a quote", "make this a numbered list", "add a divider or table", "change the caption", "save it as a draft and check it saved", "open my saved draft", or "the edit timed out, put it in again" about a post that is already in the datalab.tools Naver writing window. Not for writing a new draft from a topic, for publishing, for sidebar widget HTML, or for claims about SEO effect or how the published post looks.

Review a Naver blog post for search-facing quality - title, keyword placement, length, structure, headings, answer-style sections and internal links - using the counts and flags the datalab.tools extension returns. Use when the user asks to check a post's SEO, improve the title for search, review the outline, or asks what to fix before publishing. Not for predicting rank or traffic, for diagnosing past traffic changes, or for writing the post itself.

Check Naver blog pumasi (reciprocal engagement) targets with the user's real account - whether a post was liked, neighbour status, who reacted or commented - and, only when the user explicitly asks, open the comment box or neighbour window and fill in the text the user approved, or cancel a mutual-neighbour request the user already sent. Use when the user asks whether they liked someone's post, what their neighbour status is, who reacted to a post, or asks to prepare a comment or a mutual-neighbour request or to cancel one. Not for pressing like, for posting a comment or sending a request (the last click stays with the user), or for judging what another blogger intends.

Verified onCodex

Review a blog draft from the reader's side using the user's measured Naver blog reader data - build a profile from measured values only, then check the draft against the questions people actually ask about the topic. Use when the user asks for a reader simulation, to see a draft from the reader's point of view, or to check which expected questions a draft answers. Not for predicting how a post will perform, playing an invented persona, or rewriting the draft; the result is a thinking aid, not evidence of real reactions.

Write a research brief that cites only pages actually opened and read, never search result titles or snippets, with each fact linked to the page title and URL and no reliability scores. Use when the user asks to research a topic ("research this and sum it up with sources"), wants a sourced report, or hands over search results and page reads to turn into a brief, including when some pages could not be opened or were cut. Not for scoring source credibility, percent relevance figures, or claims about pages nobody read.

Verified onCodex

Read what Naver search shows for a query beyond blog posts - the integrated results page, cafe, Knowledge iN, video, image and Powerlink ad results - and check where a blog post ranks or what the top posts cover. Use when the user asks what shows up for a keyword, what people ask or say about it, who is advertising on it, whether my post appears for a keyword, or how the top posts are structured. Not for guaranteeing a rank, tracking rank over time, or inventing positions that no tool returned.

Analyse Naver Shopping demand - category and keyword click trends, rising keywords, season onset and shopper gender, age and device - without turning relative click indexes into sales. Use when the user asks whether demand for a product is rising, when its season starts, or which shoppers to target on Naver Shopping. Not for sales volume, revenue or buyer-count estimates, or for predicting an individual's purchase.

Verified onCodex

Rewrite one long blog post as copy-ready promotional drafts for X, Threads, Instagram and LinkedIn, each fitted to that platform's length limit and tone, with counted characters and hashtags. Use when the user wants a published post or the draft in the editor turned into social media promotion copy. Not for engagement or reach predictions, best posting times, picking a winning version, or posting.

Verified onCodex

Reconcile supplied sponsorship deadlines and approval messages into a source-linked timeline. Use when relative dates, revision requests or approval states conflict. Not for scheduling, sending, publishing or legal deadline advice.

Prepare a creator sponsorship quote from supplied deliverables, production costs and usage-right terms. Use when pricing a sponsor proposal or negotiating added scope. Not for market-rate claims, legal advice, performance guarantees or sending an offer.

Reconcile supplied sponsorship agreements, acceptance records and payment receipts. Use when checking promised fees against evidenced receipts and unresolved due conditions. Not for tax or legal judgments, bank actions or collection execution.

Analyse the tone and manner of a writer's posts and produce a style spec concrete enough to write the next post in the same voice - counted sentence features and quoted examples, no formality scores. Use when the user asks to analyse writing style, voice or tone, to capture their own blog's way of writing, to compare a competitor blog's style, or to "write like this person". Not for similarity percentages, publishing in that style, or judging a style from one post as if it were settled.

Write one beginner-friendly tutorial blog post in a problem, solution, practice structure with one analogy, inside a 12,000-character ceiling, and place it in the Naver editor when one is open. Use when the user asks for an easy explainer post, a beginner's guide, an explanation by analogy, or a hard concept in a friendly tone. Not for multi-chapter courses, minimum-length padding, self-assessment scores, badges, or quotes attributed to famous people.

Verified onCodex

Turn a topic or idea into a scene-by-scene short-form script and build it in the datalab.tools video editor - subtitles written for the ear, colour or gallery backgrounds, the user's own clips and music, paid AI narration only after the user confirms the count, and a preview and exported file when asked. Use when the user asks for a shorts or reels script, a scene plan, subtitles, AI narration or an exported video in the datalab.tools video editor, or to put a video file or music file from their computer into it as a scene background or background music. Not for camera shoots (shot lists, angles, lighting), outside video generators, or view and engagement predictions.

Verified onCodex

Save a Claude artifact used as a design mockup - a plain HTML page, a Design canvas of artboards or a Slides deck - as files outside Claude - a static HTML snapshot (plus an interactive copy), PNG or WebP images at a chosen pixel ratio and widths, layered vector SVG with real text (one layer per artboard, slide and page region, readable as layers in Inkscape and as named objects in Illustrator) and vector PDF, and the canvas data (canvas.json or deck.json, artboard sources, speaker notes, a whole-canvas overview) - by downloading the artifact's own files and rendering them in headless Chromium, with a manifest that names what rendered, what the vector output dropped and what was only approximated. Use when the user wants an artifact mockup, landing page, canvas or deck exported to HTML, PNG, WebP, SVG, PDF or JSON for a handoff, a design tool, a review, an archive or a repository. Not for designing the mockup, building production code from it, or the deck's PowerPoint export, which the Slides app does itself.

Verified onCodex

Get a design confirmed before any code - publish the proposal as a shareable document or a design canvas with one board per screen and a plan board, lead with the conclusion (tap counts, measurements, a checklist of decisions with the recommended option), upload key screenshots, label sample data, state runtime limits as dated measured facts, cite real comparable work with links, and record each answer as decided on the plan board before implementing. Use when a design or product proposal needs a reviewer's sign-off, or when a reviewer has just answered or changed a decision. Not for producing the design itself or for code review.

Review a product surface as a static HTML specimen before any code - cover every flow the plan names, build on the project's own UI framework and tokens, keep a frozen shared design frame and vary only the canvas and product mark, build one screens.html with a state hash per screen, shoot it headless at 390 and 1440 wide in light and dark, compare it with the reference product, gate on measured horizontal scroll, tap targets, contrast and overlaps, and run at least three critique rounds that also fix the checker. Use when a new product screen or flow needs a design before implementation, or when screenshots must decide whether a design is ready. Not for auditing an already shipped UI or for writing production components.

After code is generated or written from a design file or design-tool connector, compare the built values (colour, type size, spacing, radius, component used) with the design's values per component in a table with expected, built and token name, flag hard-coded values where a token exists and substitute components, and stop on any node whose design data did not load instead of guessing styles. Use when a developer or agent has built UI from a design and asks if it matches. Not for planning a build from a wireframe (prototype-from-wireframe) or reviewing a screenshot of finished assets.

Verified onCodex

Implement an approved product journey frontend-first - shared frame components, a port interface with a mock adapter and a server adapter tested against a fake fetch, a SERVER_WIRED flag, a job API that plans an outline then parts and retries only failed parts with an idempotency key and handled 402, 409 and 429 replies, versions with conflict replies, a confirm step checked on the server, receipt-only export when there is no upload service, and an end-to-end test that counts taps and text inputs at 390px. Use when building a multi-screen product flow whose backend is not ready or is built in parallel. Not for designing the screens or for a single API endpoint.

Write developer handoff notes as a state matrix per component, marking each state (default, hover, focus, empty, loading, error, disabled, long text, overflow, breakpoints) as specified, missing or not applicable from the design material given, and turning every missing state into a question for the designer instead of inventing behaviour or copy. Use when a designer or product owner asks for handoff notes or a spec from designs, screens or a component list. Not for journey gaps at the prototype stage (prototype-from-wireframe) or for building the components.

Verified onClaude Code

Plan the move from a wireframe to a prototype in a design system - map every wireframe element to a named component of the system, flag elements with no matching component instead of inventing one, keep the system's tokens for colour, spacing and type rather than the wireframe's raw values, and check the prototype against the user journey. Use when someone pastes a wireframe element list and the design system's component and token names and wants a build plan. Not for generating images or the prototype itself, for visual critique of a shipped screen, or for changing the design system.

Verified onClaude Code

Check UI colour pairs across component states (default, hover, disabled, focus ring, placeholder, input borders and other non-text elements) by computing WCAG contrast ratios in code, applying the 4.5 text, 3.0 large-text and 3.0 non-text thresholds, and reporting each ratio to two decimals with pass or fail. Use when someone lists foreground and background colours for UI states and asks if they pass accessibility contrast. Not for building a brand palette or tokens (brand-token-kit already checks contrast for the brand palette) or for a full accessibility audit.

Verified onCodexGemini CLIGrok CLIAntigravity

Prepare a product registration where options, variants and price exceptions drive the output - build the full option matrix from the company's option table, give each combination its price, stock and SKU from the stated rules, flag every combination with no price or stock rule, and never invent a price. Use when the user asks to register or list a product with sizes, colours, bundles or add-ons on a store or marketplace, to build an option or variant sheet, or to check option pricing before upload. Not for writing the product description copy or for store search optimisation.

Verified onClaude Code

Run a Discord community server deliberately - channel and role structure that matches what members come to do, onboarding, forum and FAQ upkeep that cannot go stale, moderation rules with consistent actions, announcements that are read, and bots with the least permissions. Use when setting up or reorganising a Discord server, writing its rules or onboarding, handling a moderation case, or adding a bot. Not for building a Discord bot's code.

Verified onClaude Code

Smoke-test a site or small web app on its real hosting target before launch, not only locally - upload paths, deep-link routing after refresh, form saves and where they land, secrets and config not served, no unauthenticated write path, business facts and redirects against the source of truth, analytics firing only after consent with conversions counted once, and third-party API gaps - then triage findings into launch blockers and fixes. Use when a site or app built locally, often with a generator, is about to go live or has just been deployed and needs a go or no-go check. Not for comparing an old and a new site, scanning a repository for leaked secrets, or writing page content.

Verified onClaude Code

Write the email that announces a new product, feature or store opening - one subject line and preview text that state what is new and why it matters to this reader, the change shown before it is explained, one call to action, segments for customers and prospects, the legal footer for marketing email, and a test plan for subject lines and send time. Use when the user is announcing a launch, an opening or a major update by email. Not for a recurring newsletter issue or a transactional receipt.

Verified onClaude Code

Plan and write a newsletter issue that readers open and finish - one reason this issue exists, a subject line and preview text that state it, a lead item with the most value, short sections each with one link or action, consistent recurring segments, and a check of links and rendering before sending. Use when the user is writing or planning an email newsletter or a community digest. Not for transactional email.

Verified onClaude Code

Design and test the editing path of a site that a non-developer must update - name the editors and their three most frequent changes, hold content in data a form or sheet can edit instead of in templates, require login for the editing page, run the real edit end to end, and write the editor's one-line how-to. Use when an owner, staff member or couple must change a menu, guest list, posts or pages without a developer or an AI agent. Not for building the first version of a page from a design or for payment and account systems.

Verified onClaude Code

Get a new or changed site in front of search engines without a browser session - prove domain ownership with a DNS TXT record through the Site Verification API, add the Search Console property and submit the sitemap as a service account, notify Naver, Bing and other engines through IndexNow, and wire the resubmit into the publish step. Use when the user wants a site indexed, asks to automate sitemap or Search Console submission, or launches pages on a new domain. Not for writing page copy or keyword research.

Verified onCodex

Write a regular expression that segments search queries in a search analytics tool, after naming the tool's regex dialect, and test it against a table of at least ten sample queries with expected include and exclude outcomes plus the false-positive and false-negative risks. Use when asked for a query filter regex for Search Console or a similar report, such as a brand versus non-brand split. Not for page-level URL rewrites or server redirect rules.

Verified onClaude Code

Turn a site-audit export, a list of pages or a sample of search results into a fix plan grouped by root cause, with affected URL counts, internal link placements built only from pages and anchors that exist, a staged rollout for large page sets, and a verification step per fix. Use when you hold a crawl or audit export, a page list to interlink, or a SERP sample to shape a landing page outline. Not for writing robots.txt or sitemaps (static-site-crawl-files) or submitting to search engines (search-engine-submission).

Verified onClaude Code

Check a generated landing page, portfolio or data explorer against its source material - list every claim, feature and number on the page, tie each to a source line or screenshot, strike invented claims and filler, recompute totals from the raw data, and click every call to action and download link. Use when a page was generated or heavily rewritten from a README, resume, screenshots or a dataset and must show only what the source supports. Not for site-level smoke tests of routing and consent (hosted-site-prelaunch-smoke) or launch email copy (launch-announcement-email).

Verified onClaude Code

Plan a month of social media posts a small team can actually ship - a few content pillars tied to what the business sells, a posting cadence matched to the team's time, a dated calendar with format, hook and call to action per post, reuse of one idea across formats, and a weekly check of what worked. Use when the user wants a content calendar, a posting plan or a month of ideas for Instagram, TikTok, YouTube Shorts, Threads or a Naver blog. Not for writing one carousel or one caption in detail.

Write the crawl files of a static marketing or docs site per origin and per locale - robots.txt, a sitemap.xml with canonical URLs and honest lastmod, the IndexNow key file, llms.txt - plus reciprocal hreflang for language twins and noindex headers for app documents, then fetch each file to prove it is served. Use when launching or auditing a static site's robots.txt, sitemap, llms.txt, hreflang or an English twin of a page. Not for submitting a sitemap to search engines or writing page copy.

Verified onClaude Code

Improve an app, extension or plugin store listing - the name and short description searchers see, the first screenshots, the long description's opening, and the keywords - for the people searching that store, without claims the product cannot back or keyword stuffing the store penalises. Use when writing or revising a Chrome Web Store, app store, plugin marketplace or MCP directory listing, or when installs from search are low. Not for paid ads or landing pages.

Review the output of AI photo edits - object or distraction removal, denoise, subject masking, upscaling, generated backdrops, set extension, colorway or label changes - against the original before delivery or print, at 100 percent and at print size, and write a defect list with locations and a fix or re-run for each, leaving approval to the photographer. Use when a photo has been edited with AI removal, noise reduction, masks, upscaling or compositing and needs a quality check, or when an explanatory image must not change the facts of its source. Not for generating images, batch resizing or choosing a style.

Verified onClaude Code

Check a ComfyUI graph against what is actually installed before it is written or run - take the live node list, model files and input types, use only what exists, report every missing node or model instead of silently swapping it, verify each link by output and input type, keep a condition-to-output table for batches, and map a failure log to one cause class with a before and after link diff. Use when a user asks for a ComfyUI workflow or batch, a graph references nodes, LoRAs, ControlNets or checkpoints that may not be installed, a run fails with a missing input or node error, or a custom node must extend an existing one. Not for generating the images, judging finished image quality, or other node editors.

Verified onClaude Code

Check that a design handed back as an editable layered file is real - opened in the target format with layers, editable text and independently movable elements, a download that exists - and check a bulk artboard script on a copy, per artboard, for linked state, aspect ratio, centre and margin. Use when a user asked more than once for an editable PSD, SVG or layered file and received dead links, a renamed flat image or an unfinished upload, when a delivered file or a script report on 100 or more artboards must be accepted or rejected (artboard count, linked logo, margin, centre), or when a flat image is split into objects and must recompose to the original. Not for generating the design or images, or for checking dimensions and formats of a single asset.

Verified onClaude Code

Accept or reject an AI-generated image asset against a hard delivery spec by measuring it - turn the brief into numeric checks (allowed palette, real alpha channel, aspect ratio and print size, tile seam, viewpoint, part and prop counts), measure with a script instead of judging by eye, reject baked-in checkerboards, mockup surroundings, placeholder text and off-palette pixels, then write one corrective prompt line per failure and stop after three failed rounds. Use when a generated logo, print card, texture, map, transparent graphic or limited-color design must meet a spec before it is accepted, or the user asks whether a generated image can be accepted or delivered and has a pixel report or measurements. Not for creating the image, retouching a photo or making the image look nicer.

Post-process images in batches with scripts - crop to target ratios around the subject, resize for each destination, convert formats, compress to size limits, add watermarks, and strip EXIF location and device data - keeping originals untouched and checking every output. Use when the user needs a folder of photos resized, cropped, converted, watermarked or cleaned of metadata for a website, marketplace or social platform. Not for retouching or generating images.

Verified onCodex

Make a blog or social share image (Open Graph, 1200x630) from an HTML template rendered by a headless browser, with Korean text that wraps, fonts that load, a safe zone for crops, and the og meta tags that point at it. Use when the user wants share thumbnails, OG images or blog cover images generated from a title, or when a shared link's preview image is missing, wrong or stale. Not for photo editing or AI image generation.

Verified onClaude Code

Direct AI product photos that keep the real product - start from a photo of the actual product and edit it into a scene instead of generating it from a description, record every prompt, model, seed and variant in a shot ledger, check each candidate against the reference zoomed in on logos, label text and numbers, and export only approved shots with their licence. Use when making product photos, lifestyle shots or marketplace images of a real product with an image generation or editing tool. Not for concept art of products that do not exist yet.

Verified onClaude Code

Audit a generated 2D character sprite sheet, direction atlas, equipment overlay or animation sheet against its frame plan and rig - write the frame manifest (action, phase, direction, hand) first, then check each frame for missing or duplicated directions, repeated poses, a held prop on the wrong hand, part pivots and sizes against the rig coordinates, and equipment position and timing over base frames, and return one fail-reason row per frame. Use when generated character parts, an 8-direction atlas, a weapon or equipment layer or an animation sheet must fit an existing rig or frame plan, and the user gives a frame list, a text report or measurements. Not for 3D assets, creating the art, or judging style.

Verified onClaude Code

Check a background-removed subject or cutout before it is placed on a new background - measure that the file has a real alpha channel with the share of partial-alpha pixels, find a checkerboard baked into the colour channels, require a graded matte for glass, smoke, hair and other translucent parts, composite over a light, a dark and the target background to list fringes, halos, lost thin parts and damaged lettering by location, and for frame sequences check count, order, mask polarity and edge flicker. Use when a user has an AI or tool background removal result, wants a product placed on a transparent or new background, or asks whether a cutout is ready. Not for generating the cutout or the new background, and not for checking a flat-colour logo against a palette spec.

Verified onClaude Code

Draw icons, small illustrations and diagrams as hand-written SVG - a consistent grid, stroke and corner style across a set, currentColor for theming, accessible titles, optimised output - and check them rendered at the sizes they will be used. Use when the user needs an icon set, a simple illustration, a pixel-art sprite or a diagram drawn as SVG code. Not for photos or AI-generated images.

Verified onCodex

Check romanized access points for catalog records of non-Latin material (Cyrillic, Greek, Arabic and similar) - the transliteration standard named and applied consistently, original script kept in its own field beside the romanized form and a labelled gloss, unreadable or ambiguous characters flagged instead of guessed, and searchable variant spellings listed. Use when cataloguing or describing books, titles or names in another script so users can find them, including from OCR of a title page. Not for translating running text (translation-postedit-qc) and not a substitute for a cataloguer's authority-file check.

Verified onClaude Code

Prepare and check the translation handoff for a licensed comic or webtoon from the page data the user supplies - a page, panel, bubble, speaker and reading-order table built first (right to left or top to bottom) and given to the translator as context, each translated line checked against its bubble for length and line breaks with overflow listed, complex panels and off-panel or monologue lines flagged for manual confirmation, and webtoon slices compared with the original scroll for order, seams and height. Use when OCR lines, panel boxes or translated text for manga, comics or a long-strip webtoon must be ordered, fitted or exported. Not for subtitles (subtitle-qc), translating text itself, or editing artwork.

Verified onClaude Code

Compare gettext PO/MO message identity, typed format arguments and plural selection with supplied consumer rendering controls. Use when compilation succeeds but localized runtime output or branch delivery is uncertain. Not for prose translation, XLIFF merge reconstruction or literal format-string identity enforcement.

Verified onCodex

Check generated text against a language policy the user set - regional variety (Brazilian vs European Portuguese, a spoken Arabic dialect vs the standard form), speech level (Korean haeyo vs banmal, formal vs informal address) and the foreign-language lines a character is meant to speak - span by span and turn by turn, listing every drift by line with its reason and fixing only the drift. Use when a reply drifted out of the requested dialect, variety or politeness level, mixed speech levels, slipped in unrequested foreign words, replaced a requested foreign line with a description, or the user has to repeat the instruction on later turns. Not for checking translation fidelity against a source text or for producing the first draft.

Verified onClaude Code

Check subtitle cues or a dubbing-plus-subtitle package against limits - characters per line, lines per cue, reading speed, minimum duration, monotonic non-overlapping timing - and report violations by cue id, with phrase-boundary line splits. Use when creating, reviewing or revising SRT or VTT subtitles, or preparing a dubbing script and subtitle set for a video. Not for transcribing audio or translating the content.

Verified onClaude Code

Review a machine or AI translation against its source before delivery - segment-by-segment omissions and additions, footnote and numbering parity, unnatural phrasing with a fix and a reason, a term glossary, and the spots a native reviewer must confirm. Use when a user asks to check, polish or make an AI translation sound natural, especially academic text with footnotes. Not for producing the first translation from scratch.

Verified onClaude Code

Check XLIFF unit/source/target roles and permission-aware inline-code reconstruction against approved extraction and merge contracts. Use when translations overwrite source, required targets disappear or valid XML loses native placeholders. Not for linguistic quality or treating every code reorder or absent target as invalid.

Verified onCodex

Review AI-transcribed MusicXML meter and proposed bar boundaries against supplied source-audio annotations and owner-approved beats while preserving voices, chords and ties. Use when transcription produces irregular meters or a proposed rebar must be checked for lossless musical timing. Not for forcing 4/4 without beat evidence or claiming independent audio recovery from XML sums.

Audit MusicXML percussion note instrument references against score/MIDI definitions and an approved drum map, separating display position from supplied playback sound. Use when adding an instrument changes a drum sound while the staff looks the same. Not for inferring percussion identity from staff position or inventing a target sound map.

Verified onCodex

Compare MusicXML repeat and jump navigation with an approved measure-visit sequence and supplied playback evidence, including context-dependent Fine termination. Use when D.C., D.S., Coda or repeat symbols survive import but playback skips or stops incorrectly. Not for score composition or declaring playback correct from schema or glyphs alone.

Verified onCodex

Check MusicXML written, sounding and linked tablature pitch against an explicit transpose and octave reference across import and export. Use when notation looks unchanged but a linked tab staff or sounding octave changes after a roundtrip. Not for inferring transposition from instrument names or judging audio that was not observed.

Audit a search or shopping ad account from exported data - state the date range, currency, conversion definition and attribution window first, compute budget pacing against the prorated budget for the elapsed days, rank waste patterns with spend and sample size, check feed items against required attributes, and keep evidence apart from untested hypotheses. Use when someone shares an export and asks for an audit, a pacing check, keyword gaps or duplicate keywords, or a feed review. Not for changing the account, forecasting results, guaranteeing savings, or platform policy rulings.

Verified onClaude Code

Reconcile the conversions an ad platform reports with the leads or outcomes in a CRM or backend export - pick the join key and report the match rate first, clean the CRM count of duplicates and test records with the arithmetic shown, name the attribution window and date-basis differences between the systems, report counts per funnel stage and the gap in absolute terms, say which part of the gap is explained and which is not, and list the tag, consent and event checks that need a live browser or tag debugger as not performed. Use when someone compares ad conversions with CRM, backend or order data, or asks why the two numbers differ. Not for changing tracking, assigning credit between channels, personal-level matching, or an audit of the ad account itself (use ad-account-audit-readout).

Verified onClaude Code

Review a pasted search-term report, keyword bid table or hourly export for an advertiser - classify terms against the stated business constraints into negative candidates with the reason without ever excluding a term that converted, propose manual-campaign promotions only for terms that meet a stated conversion count and window, compute ACoS per row against the target and margin, mark rows with too few clicks as not decidable, and give dayparting as scenarios with the sample caveat. Use when someone shares search terms, bids or hourly performance and asks for negatives, promotion candidates, bid changes or a schedule. Not for changing the account, forecasting results, or whole-account audits.

Verified onClaude Code

Plan and verify a batch grade, film look or RAW-development tool from the settings and logs the photographer supplies - white, skin and exposure continuity compared per scene with intended lighting differences kept, preview checked against export for colour space, clipping and resolution, input files mapped one to one to outputs with unopened files listed, and style approval left to the photographer on a sample. Use when a look is applied to a batch, a tool is built for non-standard RAW or DNG files, or preview and export differ. Not for culling and not for generating images.

Verified onClaude Code

Plan and review a photo cull from the scores, flags or file lists the photographer supplies - criteria set per subject (wildlife, event, burst), calibrated on frames the photographer already keeps, technical flaws kept apart from moments that matter, rank-and-flag only with nothing deleted, and groups or frames missing from the report listed. Use when an assistant or app rates, rejects or picks best frames from wildlife, event or burst sets. Not for grading or exporting a batch and not for generating images.

Verified onClaude Code

Draft and check titles, descriptions, keywords and upload CSVs for an existing photo batch against the destination's limits - field lengths, keyword syntax and count, ordering, escaping of commas and quotes, encoding, one row per exact filename - keep unknown locations, identities and species out of the text until the photographer confirms them, and list files with no notes. Use when a photographer asks an assistant for stock-site metadata or a CSV for a batch. Not for generating or editing images, choosing keepers, release or licence validity, or sales promises.

Verified onCodex

Compile a media coverage or clippings report only from what was pasted - one row per item with outlet, date, headline and URL copied from the pasted links or export, the message pulled out of the pasted text, metrics taken only from the supplied tool export with the field named, items missing a date, URL or export row marked, date or outlet disagreements between sources listed, no extra clippings added, nothing fetched, and no advertising-value-equivalent figure computed. Use when a PR or comms person asks for a coverage report, a clippings summary or total reach from pasted links and an export. Not for sentiment or reputation analysis, for finding coverage, or for press release approval (press-release-fact-and-quote-approval).

Prepare a spokesperson for an interview with a briefing pack - likely questions including hostile ones, the approved key message for each, a supporting figure with its source and date taken only from pasted material, a staleness flag for figures older than 12 months or older than the interview date allows, bridge-back lines, and an off-limits list kept separate. Use when someone asks for interview prep, a Q and A sheet or talking points for a named spokesperson and supplies source material. Not for writing answers that use figures nobody approved, for media lists (media-target-and-angle-verification) or for press releases (press-release-fact-and-quote-approval).

Verified onCodex

Prepare a media target list and pitch angles for a PR pitch without inventing facts - each journalist candidate gets a beat, a byline URL and the date it was seen taken only from pages the user pasted, or is marked unverified; no email or phone that is not on a source page; each editorial-calendar or news hook is tied to approved expertise only, with its event date and source, and a passed deadline is not treated as open; sensitive or fast-moving news is flagged for a human check. Use when someone asks to find journalists, outlets, editorial-calendar slots or news hooks for a client pitch. Not for writing the pitch itself or judging a named person.

Verified onClaude Code

Draft a press release or statement in which every sentence carries the id of an approved fact, quotes are copied byte for byte from text the named spokesperson approved, and claims such as AI-powered, first, leading or record-setting are flagged for substantiation instead of written as fact. Sentences with no approved fact are removed and listed, and the draft ends with an approval checklist. Use when someone asks for a press release, announcement or public statement and supplies a fact sheet, approved quotes or both. Not for media lists or pitch angles (media-target-and-angle-verification) or legal review of claims.

Verified onCodexGemini CLIGrok CLIAntigravity

Edit an existing slide deck without losing content and keep hand fixes alive across regeneration - record a per-slide preservation list of required text and objects before editing, diff the file against it afterwards and report every removed, moved or changed element, keep content source and layout template apart, write manual fixes back to the source, and warn which hand edits a regeneration would overwrite. Use when an AI edit of a pptx or HTML deck dropped a table, image or text, or small manual fixes vanished when the deck was regenerated. Not for building a deck from a template (use pptx-template-build) or reviewing numbers in a finished deck (use deck-qa-review).

Verified onClaude Code

Check a finished deck before it is presented or sent - numbers that disagree between slides, charts that do not match their labels or data, totals that do not add up, dates and weekdays that do not match, text overflow, inconsistent fonts and terms, and claims without a source. Use when the user asks to review, proofread or QA a slide deck or a pptx or PDF presentation. Not for restructuring the storyline.

Verified onCodex

Build the storyline of a presentation before any slide is designed - the one answer the deck argues, the supporting points grouped under it, and one full-sentence message per slide that reads as the whole story on its own. Use when the user asks for a deck outline, a presentation structure, slide titles or a pitch storyline. Not for designing slides or building a pptx file.

Verified onClaude Code

Build a PowerPoint file that respects a template - slides created from the template's layouts and placeholders, the slide master's fonts and colors, native charts and tables that stay editable, and checks for overflow and missing placeholders. Use when generating or editing a .pptx from data or an outline with a company template. Not for designing a template from scratch or writing the storyline.

Verified onClaude Code

Rehearse the hard questions an executive or reviewer will ask about a plan before it is presented - list the questions by what each one attacks, attach the evidence that answers it, mark which answers rest on data and which on assumption, flag logic gaps in the plan, and draft short answers. Use when someone has a plan, proposal or deck outline and a goal and wants to prepare for questions from leadership. Not for checking a finished deck's numbers (use deck-qa-review) or building the storyline (use deck-storyline).

Verified onClaude Code

Check a book that goes to print and ebook before the edition is fixed - diff the typeset text against the approved manuscript and tie each difference to page and edition, check print (trim, margins, bleed, running heads) and EPUB (headings, table of contents, reflow, alt text, metadata) against separate rules, and pair automated EPUB accessibility results with the manual checks they cannot prove, never calling a file conformant from an automated pass. Use when an author or small press pastes typeset text, layout facts, an EPUB manifest or checker output and asks whether the proof or the files are ready. Not for writing or editing the manuscript, designing the layout, legal accessibility sign-off, or generating the files.

Verified onClaude Code

Copyedit a book-length manuscript through a project style sheet - name the style manual and edition, count spelling, number, capitalisation, name and hyphenation variants in code, settle one choice per item with a rule id or a marked house choice, apply the sheet chapter by chapter with a change log (location, before, after, rule), leave the author's voice and singular they alone, and re-run to prove no variants remain. Use when a long manuscript is copyedited and the same decision must hold across chapters, or when an earlier AI pass fixed some instances and missed others. Not for one client's marketing rulebook (client-rulebook-copy-check) or story continuity (manuscript-continuity-audit).

Verified onCodexGemini CLIGrok CLIAntigravity

Check approved revision choices, comment remnants and exact delivered DOCX identity. Use when No Markup or zero visible revisions is offered as proof of a clean review copy, or recipients see old exchanges. Not for deciding legal edits or blindly accepting every change.

Check a selected EPUB package's manifest, spine and local resource references against packaged files and supplied reader evidence. Use when previews or navigation checks hide missing publication resources. Not for prose editing, general TOC review or accessibility certification.

Check a book index, made by hand or with AI, against page-numbered final proofs - parse each entry and locator, search the term and its variants on the cited page in code, mark every locator found, not found or off by one page, check that each page range holds the term throughout or is queried, and check that every see and see also target exists as an entry. Use when an index is checked before release against the final page proofs, or when an AI-made index may cite pages where the term never appears. Not for choosing index terms or for layout checks (book-layout-proof-and-epub-check).

Verified onCodex

Audit new chapters or scenes of a long fiction manuscript against the author's own canon ledger, outline, character notes and chosen tense and point of view - setting contradictions, outline drift, repeated information, unmotivated actions and tense or POV slips - reported as a table with the earlier line quoted and every judgement left to the author. Use when a writer pastes a draft chapter with a ledger, outline or notes and asks for a continuity check, a consistency pass or an outline comparison. Not for rewriting prose, grading style or genre taste, or proofreading spelling.

Verified onClaude Code

Compress a long series bible or season outline to a short treatment while keeping each episode's cause, choice and consequence chain, tie each major event to the lead's goal, choice and cost, mark steps where motivation exists only in the script and not on the page, and list what was cut and why with names and dates kept from the source. Use when a writer must shorten a bible to a page limit or a producer says the character's change or motivation is not visible in the outline. Not for line edits, new plot, or a continuity check of a draft chapter (manuscript-continuity-audit).

Verified onClaude Code

Make Korean card news for Naver blog posts and Naver channels - square or portrait cards with large Korean headlines that wrap at word boundaries, one message per card, readable at phone size, the post text around the cards for search, and alt text. Use when the user wants Korean card news (a series of text-led square images) for a Naver blog or a Korean social channel. Not for English carousels or video thumbnails.

Write the slide-by-slide copy of an Instagram or LinkedIn carousel - a hook, the problem, the value, a summary and one call to action - with one idea and a short line per slide. Use when the user asks for carousel copy, card news text or a swipe post outline. Not for designing the slides or writing a blog post.

Verified onClaude Code

Render carousel slides from HTML to PNG with a headless browser at the platform's sizes (1080x1350 portrait, 1080x1080 square, 1080x1920 story), check every slide before export, and export only after the user approves. Use when the user has carousel copy or a design and wants the slide image files produced. Not for writing the copy or editing photos.

Verified onClaude Code

Check an AI-generated video clip against its brief frame by frame before regenerating - turn the request into measurable checks (locked camera, loop seam, stray text, prop era, hand continuity, clean borders, join between clips), sample frames with ffmpeg, then rewrite the prompt around the one failing constraint and cap the retries. Use when a generated clip drifts, will not loop, shows unwanted text or anachronistic props, has black borders after upscaling, or does not join the next clip, and the user keeps regenerating blindly. Not for planning footage, editing real recordings or subtitle timing.

Verified onClaude Code

Plan and check mouth animation for a non-human or cartoon character, a sung track or a long audio file by building an explicit viseme timeline and verifying it against the audio length, closed-mouth sounds, segment joins and sampled frames. Use when automatic face-based lip sync misses the character's face, mouths look like speech during a song, a render fails on long audio but works on short clips, or frames where no face is detected break the sequence. Not for subtitle timing, for generating or rendering the animation itself, or for human talking-head edits.

Verified onClaude Code

Map product demo claims and UI steps to versioned support documentation, observed operations and explicitly simulated scenes. Use when a script or app recreation may show unsupported features, confuse local and cloud modes or imply a successful operation without a receipt. Not for market research, executing the demo or certifying a product from mock screens.

Verified onCodex

Build a product explainer or release video in Remotion - compositions with fixed width, height, fps and duration, scenes driven by frame math, assets loaded before render, fonts that are licensed and loaded, a render manifest, and the rendered file verified with ffprobe. Use when Remotion compositions/rendering are selected or an existing Remotion render is missing elements, frozen, flickering or mistimed. React UI alone does not establish a Remotion video workflow. Not for plain HTML/canvas without React or Remotion, stack-neutral deterministic animation verification, or editing filmed footage.

Verify a code-driven animation (HTML or canvas, SVG, Manim, a scripted trailer) against its required event order, scene continuity, total duration, text behaviour and mechanical linkage by reading the timeline in the code and checking sampled frames. Use when a user asks for or has received a scripted animation with events that must happen in a given order, motion reused from an earlier file, a mechanism that must move in phase (feet on pedals), text that wraps or clips at scene changes, or preview timing that differs from the export. Not for generating the animation, AI video clips, or editing recorded footage.

Check whether fictional or user-owned characters in AI-generated stills or video still match their approved reference sheet across shots and episodes - split fixed traits from per-episode state, compare every re-appearance to the original reference and never to the previous shot, label each as match, intended change, unintended drift or cannot judge with frame and timecode evidence, check characters in one shot stay distinct, and leave approval to the director. Use when a character's face, hair, outfit, props or build change between generated shots, outfits reset to the first state, or two characters blend. Not for identifying or profiling a real person from a face, or for generating the footage.

Verified onClaude Code

Verify subtitle-to-media synchronization and diagnose offset, drift and edit discontinuities with independent residual checks. Use when the user requests timing verification or correction using identifiable subtitle/media versions and supplied observations or clock anchors.

Edit talking-head or screen-recorded video with ffmpeg - cut long silences and false starts without clipping words, level and clean the voice, burn in or attach subtitles, and export to the platform's format - keeping the original and a list of every cut. Use when the user wants a recorded talk, lecture, interview or community video tightened, captioned or exported. Not for animated explainers or color grading.

Verified onCodex

Choose clips or build a paper edit from a timed transcript so every cut lands on a complete thought and a word boundary, and check each in and out point before anything is rendered - read the line before and after, add handles, keep breaths and natural pauses, count the filler removed, and list the timeline with quoted first and last words. Use when someone asks for a short clip from a podcast or interview transcript, a transcript-based rough cut, or filler and pause removal that must still sound natural. Not for generating video, rendering the cut with ffmpeg, subtitle timing or audio mixing.

Verified onClaude Code

Create audio-timed Korean captions or screen cues with spoken/display mapping and SRT, WebVTT and frame-numbered exports. Use when the user requests alignment of supplied known narration text with an identifiable voice track.

Turn a video request into a brief and a scene-by-scene storyboard that a renderer or editor can execute - goal, audience, platform, length, resolution and frame rate, then each scene's duration, visuals, on-screen text, voice-over and asset sources, with acceptance criteria. Use when the user asks for a video plan, storyboard, shot list or brief for a short-form, explainer or product video. Not for rendering or editing the video.

Compare approved personal or literal creator identities, order and roles across bibliography source metadata, export, consumer parsing and rendered style. Use when Zotero, BibTeX or CSL output appears to split, reorder or change multilingual authors or institutional names. Not for DOI existence, citation renumbering, guessing family names from strings or global brace rewrites.

Check AI-assisted wet-lab work before use - redo every dilution and unit conversion with units carried through, compare generated results tables with the raw numbers at their stated precision, and test whether a cited paper or kit manual actually contains the requested method, sample type and condition. Use when an assistant gave a dilution, a protocol parameter with a citation, a rounded results table, or says a paper contains a specific method. Not for designing experiments, replacing a pilot test, finding substitute citations, safety sign-off, or checking a structure entry for a ligand (use structure-record-ligand-presence-check).

Verified onCodex

Compare OME microscopy physical quantities and series/IFD-to-ZCT plane identities through export and consumer reimport observations. Use when pixels look unchanged but scale, time interval, channels or plane ordering change. Not for biological interpretation, OCR alignment, web-image resizing or guessing calibration from screenshots or DPI.

Check a re-implementation or port of a published method before it enters a workflow - the paper's equations and parameters as a numbered checklist, the authors' reference code and data located and run first with version and licence recorded, one published number reproduced before any adaptation, every deviation listed, and the ambiguous parts of the method text named. Use when asked to implement an algorithm from a paper or to extract and adapt a paper's published code. Not for checking a paper's citations (research-source-verification) and not for judging whether the method is scientifically sound.

Verified onClaude Code

Renumber a numbered reference list and keep every in-text citation in sync - an explicit old-to-new mapping built before the text is touched, single, grouped and range citations rewritten and re-compressed, citations to deleted entries flagged instead of dropped, and uncited or unresolved references listed. Use when references are reordered, inserted or deleted in a manuscript with numeric citations such as [3], [2,7] or [4-6]. Not for author-year styles, citation formatting or checking that references exist (research-source-verification).

Verified onClaude Code

Audit a research answer on a field, policy, history or genre for missing key items and wrong dates - fix scope and period, build the item list from at least two independent sources, date and source every entry in order, separate proposed from enacted and claims from findings, and report coverage gaps and searches not run. Use when a survey, timeline or background report looks incomplete or mixes proposals with enacted rules. Not for checking that individual citations exist; use research-source-verification for that.

Verified onClaude Code

Check AI-suggested papers, citations, measurement instruments and extracted statistics before they go into research notes or a manuscript - match each citation to a real index entry, pull instruments and scales only as quoted text, and screen papers against the research question. Use when an assistant recommends a paper, a scale or a reference list, or when you extract methods or measures from a paper. Not for formatting citations or writing the literature review itself.

Verified onClaude Code

Check a recommended structure entry against the record text the researcher pastes - restate the target, ligand name or code and any ID given, search the pasted record for that ligand code, chain and occupancy, report "not found in the pasted text" when absent, never name an entry ID from memory, and list what to confirm in the source database. Use when an assistant or colleague recommends a structure for a named ligand or target and the claim must be checked before use. Not for docking, scoring or judging binding, and not for other citation checks (use lab-calculation-and-claim-check).

Verified onCodexGemini CLIGrok CLIAntigravity

Plan and accept a bioinformatics refactor that must keep results equal - replacing a hand-written analysis step with an official package, porting a shell or script pipeline to a workflow engine, or accelerating it on GPU or in parallel - by freezing a baseline with seeds, versions and checksums, mapping every old step to its replacement, comparing outputs at a stated tolerance and reporting explained and unexplained differences apart. Use when a researcher asks for a migration plan or acceptance check for an analysis that already produced results. Not for designing a new analysis, choosing statistics for new data or SQL migrations.

Verified onClaude Code

Write or review a differential-expression script, a GTF or GFF annotation parser or a publication figure script so that input contracts, annotation traps (1-based inclusive coordinates, gene id version suffixes, several transcripts per gene, quoted attributes), filter counts and a hand-checkable fixture are stated before the code is trusted; scripts are reviewed and the fixture's expected numbers are worked by hand, never presented as run. Use when someone asks for such a script or asks whether one is right. Not for refactoring a finished pipeline for parity, choosing statistics for new data or interpreting results biologically.

Verified onClaude Code

Compare ordered FASTQ mate identities and eligible read conservation against source name/category inventories and explicit singleton and catch-all policies. Use when paired exports have equal counts but wrong pairing, orphan reads disappear or BAM-to-FASTQ recovery is questioned. Not for alignment, biological interpretation or reconstructing unavailable mates or clipped bases.

Verified onCodex

Audit an AI-proposed or newly adopted omics analysis (single-cell clustering, differential testing, marker-based cell-type labels, an unfamiliar method) for hidden assumptions, design gaps and unsupported biological claims - an assumptions table checked against the described data, replicate-unit and covariate gaps, testing scope, and a claim table where each label needs a source that was actually opened. Use when someone asks to check or review an analysis before it goes into a paper, a talk or a decision, or asks whether a method fits their data. Not for running the analysis, judging scientific validity, or replacing review by a statistician or domain biologist.

Verified onClaude Code

Compare exact reference spans and represented alternate sequences while tracing VCF allele indices, genotypes and annotations through normalization. Use when normalized variants cannot merge, REF repair is proposed or splitting changes allele-indexed fields. Not for clinical interpretation, variant calling, guessing a reference or treating allele swapping as strand repair.

Verified onCodex

Audit answers a chatbot, helpdesk or staff member already gave against the company's current rule table - split each answer into claims, tie every claim to a rule row and version, and flag dropped conditions, contradicted rules, invented permissions and claims no row supports, with corrected wording for a person to approve. Use when a transcript or answer log must be checked against a rule table after the fact. Not for deciding one case (policy-exception-handling), writing the policy, answering a new question, or sending anything to a customer.

Verified onCodex

Apply a company's versioned rule table (refund, cancellation fee, exchange, discount, compensation) and its checklist with a named exception list to one case or document, including tax, trade and administrative documents and quality-defect classification - sort the case into normal, exception or missing information, cite the rule version and row used, and hand discounts, refunds and approvals to a person instead of applying them. Use when the user asks how a policy applies to a customer case, a claim, a refund or fee request, or asks to review a document against a checklist. Not for writing the policy itself or for legal advice.

Verified onClaude Code

Test a rule table or checklist before it goes into use - parse it into rows, find overlapping and uncovered ranges, ambiguous boundaries such as at-least against over, missing effective dates and exception rows that shadow a rule, then generate one test case per boundary and per exception row with the row each one hits and list the cases that hit two rows or none. Use when a rule table, fee schedule or checklist is about to be put into use or handed to an assistant. Not for gap review of a prose policy against a checklist (policy-draft-gap-review in hr-ops-plus), for applying the table to a live case (policy-exception-handling), or for rewriting the table.

Verified onCodexGemini CLIGrok CLIAntigravity

Turn a seller's photos and visible markings of a second-hand item into a ranked list of identification candidates, each tied to the visible feature behind it, with what the photos cannot confirm (exact model, authenticity, condition, working order, value) and a short list of extra photos to ask for. Use when someone preparing a resale listing describes or shows photos and asks what the item is. Not for pricing or valuation, authenticating an item, grading condition, or writing the listing copy.

Verified onClaude Code

Rewrite many product titles and descriptions from a spec sheet or catalog export for a marketplace or shop, using only facts in each SKU's own row, with titles counted by code against the channel limit, banned words and required attributes checked, and a re-upload sheet with unchanged SKU keys plus a rejected-claims list. Use when a seller or merchant pastes spec rows and wants titles or descriptions for a batch of SKUs. Not for identifying an unknown item (item-identification-from-photos), store-page SEO, or legal review of claims.

Verified onCodex

Triage a shopping feed or marketplace diagnostics export, grouping disapproved or flagged products by the attribute the issue names, giving each fix as a value taken from the product source or marking it needs source, never inventing identifiers, brands or prices, and re-checking every row with counts per issue before and after. Use when a merchant pastes diagnostics and feed rows after items are disapproved or missing attributes. Not for ad bidding or campaign structure (paid-media-plus) or for policy and legal questions about what may be sold.

Verified onCodexGrok CLI

Set reorder points and order quantities for a small shop from a pasted sales sheet with lead time, minimum order quantity (MOQ) and pack size, with every input's unit written down and converted, the safety-stock rule stated, order quantities rounded up to MOQ and pack, scenario rows for higher demand and longer lead time, and thin-history SKUs flagged. Use when a shop owner asks when and how much to reorder from weekly or daily sales. Not for forecast backtesting on longer history with stockouts, pricing, or supplier selection.

Verified onCodex

Debug a BI measure, calculated column or conditional-format rule that returns wrong totals - name the model relationships and filter context first, reproduce the wrong value for one filter state and compute the expected value by hand, simplify only while subtotal, grand total and filtered results stay equal, derive per-group scales from the model, and treat a live master file as having no past state. Use when a Power BI, DAX or similar measure gives rows that do not add to the total, is over-nested, needs a per-group color scale, or a past-period filter on a changing master file breaks counts. Not for building a new dashboard layout or for SQL tuning.

Verified onClaude Code

Prove a generated calculator, converter or numeric tool correct against known reference cases before anyone uses it - state the formula, domain and unit definitions first, build a reference table of at least 5 cases with boundaries and out-of-range inputs, test round trips and dimensions, reject duplicate options and rounding that hides errors, and check extreme inputs. Use when an AI-generated or hand-written calculator, unit converter, pace or load tool, pricing or total calculator, or numeric engine must be trusted, or a user reports a wrong result from one. Not for defining business metrics or writing unit tests for application code.

Verified onClaude Code

Check, build or caption charts by recomputing every number from the source data instead of reading the picture - bar totals and stacked sums against printed labels, axis origin and scale, bar length ratio against value ratio, legend to series match, correlation coefficients with sample size and missing values - and test label overflow, tooltips and missing categories across widths and many charts. Use when a user asks an AI to verify, caption or compare charts or scatter plots, to say which bars do not match their labels, to judge a correlation from a plot, or when long axis labels clip in generated charts. Not for choosing chart colours or style, or for statistics with no chart.

Verified onClaude Code

Write or review ClickHouse SQL that reads the primary key, states its timezone, deduplicates correctly and takes values as typed parameters - and check the plan before calling it fast. Use when the user writes, debugs or reviews a ClickHouse query, a materialised view or a table's ORDER BY and partition key. Not for PostgreSQL or MySQL tuning.

Verified onCodex

Build and read cohort retention - users grouped by the period they started, the share still active in each later period, with the activity definition, the timezone and the incomplete latest periods stated - so retention is compared like for like instead of from a single blended number. Use when the user asks whether users come back, how retention changed after a release, or wants a retention query or chart. Not for revenue forecasting.

Verified onClaude CodeCodexGrok CLI

Check a CSV export against spreadsheet formula interpretation and machine-value contracts, including import and save/reopen evidence. Use when reviewing CSV safety, escaping or spreadsheet delivery. Not for exploit payloads or a universal sanitizer guarantee.

Verified onCodex

Design an A/B test before it starts and read it honestly after - one primary metric, the minimum effect worth detecting, the sample size and duration that follow, a stopping rule fixed in advance, guardrail metrics, and a readout that reports the interval and refuses to call a winner from peeking or from a sub-group found afterwards. Use when planning an experiment, deciding how long a test must run, or interpreting test results. Not for tests that cannot be randomised.

Verified onClaude Code

Build a Sankey, country map, heatmap or publication figure from parsed exports, handwritten records or country statistics without silent mismatches - parse the source into a table and check group totals against the export totals before drawing, join statistic keys to the real boundary or zone list and list unmatched and renamed entries with the aggregation unit, state that text-derived counts measure mentions not visits, and give rerunnable code. Use when a figure is made from tag exports, PDF tables, visitor notes or per-country figures. Not for checking an existing chart's labels or ratios, or generic plotting style.

Verified onClaude Code

Compare two lists or clean identifier, name and address columns before a hand-off - state the normalization first, report differences in both directions with counts that reconcile, list every merged or dropped key, and keep postal codes and leading zeros. Use when asked what is in one list but not the other, to combine address columns, or to clean registration, asset or test identifiers before an upload. Not for fuzzy record linkage at scale or database deduplication in SQL.

Verified onClaude Code

Recover the business logic of old spreadsheet macros or layered code and refactor shared parts without changing behavior - tabulate conditions, inputs, outputs and side effects per routine, keep every difference between near-duplicates as a parameter, treat shared state as an input, and compare outputs before and after on sample rows. Use when a user has old macros or layered code and needs the logic documented, a shared function extracted, or the layer a change must touch found. Not for rewriting in another language, performance tuning or new features.

Verified onClaude Code

Define a product or business metric so two people computing it get the same number - the entity counted, the event that counts, the time window and timezone, inclusions and exclusions, the source table, the query, and the guardrail metric that catches gaming. Use when the user introduces or questions a KPI such as active users, retention, conversion or churn, or when two dashboards disagree. Not for choosing which metric matters most.

Verified onCodex

Read Naver search and shopping trend data correctly - Data Lab ratios are relative to the peak inside one request, not search volumes; absolute monthly volumes come from the search ad keyword tool; compare keywords only within one request and say which filters were used. Use when analysing Korean search interest, comparing keywords on Naver, or turning Naver Data Lab or keyword-tool numbers into a recommendation. Not for Google Trends or web analytics.

Verified onCodexGrok CLI

Turn a set of open survey responses into deduplicated idea groups that keep every response id, merge only ideas with a stated similarity reason, report counts in, merged, unclassified and out and reconcile them to the actual number of responses, process large sets in labelled batches so late responses are not dropped, and compare theme sets from different methods (topic model, human coding, AI themes) in a match and difference table with sample quotes. Use when an analyst has many open-ended responses and wants merged ideas with counts or wants to compare theme sets from different methods. Not for recounting a finished summary table or anonymizing respondents (use survey-recount-and-anonymize), significance testing, or deciding what the themes mean for the business.

Verified onClaude Code

Read a retrieval ranking experiment from labelled queries and ranked results - calculate Recall at k and MRR, separate language slices, audit the candidate corpus and latency measurement, and reject a misleading aggregate win. Use when comparing lexical and embedding search, choosing a multilingual retriever, or asked whether better average retrieval scores justify a switch. Not for evaluating generated answers or implementing a search engine.

Verified onClaude Code

Write a spreadsheet formula or custom function for a named tool, version and locale - confirm the engine and list the functions it lacks before proposing anything, use the locale's argument separator, test blanks, text numbers and separators on sample rows, rebuild highlight-based totals from the conditional formatting rule because cell color is not a value, and give a fallback when a function is unavailable. Use when a user names the exact spreadsheet app, version or locale a formula must run in, or asks to sum or count highlighted cells. Not for comparing two lists, macros, or BI measures.

Verified onClaude Code

Before recording survey or interview results someone hands you, recount every tally from the per-row table instead of trusting the summary, name the bucket definitions each count uses, report where the summary and the recount differ, replace company or respondent names with numbers in anything shared, keep identifying data out of the repository (aggregates only), and search for the names before committing. Use when a survey summary, interview tally or research table is about to be written into a plan, deck, document or repository. Not for designing the survey or for statistical significance testing.

Verified onClaude Code

Verify actual coordinate tuple order, API axis mapping and selected CRS operation against independent forward location controls and saved-output interpretation. Use when a transform gives plausible but wrong locations, an axis default changes, or inverse roundtrip is offered as proof. Not for geometry repair, guessing missing CRS or surveying legal boundaries.

Write a QGIS or ArcGIS label expression or a PostGIS spatial view against a stated schema - quote the exact field names, types and geometry column, handle NULL and type casts explicitly, test the expression on three sample rows with the expected output worked by hand, and for views state the SRID, index use, join cardinality and the one-row-per-key check. Use when someone asks for a label like name plus area in hectares, a conditional label, or a view that joins layers by spatial relation or largest overlap. Not for cartographic style design, for performance tuning of a whole database or for raster algebra.

Verified onClaude Code

Plan the cleanup and merge of several polygon layers - run validity checks first and count error types, align CRS and field schema across layers with a mapping table, repair with a named method, compare feature count and area before and after, check boundaries for gaps and overlaps, and keep a repair log of changed feature ids. Use when polygons are invalid, layers from different regions use different CRS or field names, or they must be merged into one GeoPackage with one style. Not for raster data (see raster-nodata-and-scaling), for deciding legal boundaries or for writing the merge for the user without a validity report.

Verified onClaude Code

Plan how to get the area per slope class from contour lines - interpolate a DEM with a named method and cell size, derive slope as percent or degrees (state which), reclassify at the stated breaks, sum areas in a projected CRS, check the class sum against the study area, and state how the contour interval limits accuracy, using the bundled calculator for percent to degree and spacing. Use when a planner or student has contours and wants hectares or shares of land under given slope limits. Not for measured survey slopes, for slope stability or earthworks design, or for raster scaling (see raster-nodata-and-scaling).

Verified onClaude Code

Plan a replacement of a proprietary GIS analysis tool with a free or other licensed one - name the algorithm and every parameter of the original (observer and target height, distance, curvature and refraction, cell size, units), map each to the candidate tool's option, list the defaults that differ, and compare CRS and output on a small test area before any claim of equivalence. Use when a license lacks a tool such as a viewshed, hydrology or network analysis and someone asks for an open alternative or a QGIS or GDAL equivalent. Not for installing software, for license advice or for tools that need no equivalence check. Rests on one report.

Verified onClaude Code

Plan how to shrink, trim or aggregate a large raster without changing its meaning - separate NoData from valid zero before trimming, pick an integer type with scale and offset from the stated error tolerance with the bundled calculator, keep a NoData code outside the valid range, and compare the result with the original on one tile and on grid extent and alignment. Use when someone wants a smaller float raster, wants empty borders or zeros cut, or wants a tiled or parallel aggregation checked against a single run. Not for choosing a projection, for vector data (see gis-geometry-repair-and-validate) or for running the conversion for the user.

Verified onClaude Code

Implement or review a Lemon Squeezy integration - hosted checkout, signed webhooks, subscriptions, refunds, catalog reads and fees - and check each API detail against the official Lemon Squeezy docs, including the details that are easy to get wrong (the refund endpoint, snake_case checkout keys, no webhook delivery id, the hidden default variant, ten-row pages). Use when the user builds, debugs or reviews checkout, webhooks, refunds or subscription sync with Lemon Squeezy. Not for other payment providers.

Reconcile supplied Lemon Squeezy store, product, variant and current Price evidence against an approved catalog expectation and prepare a mode-specific mapping handoff. Use when reviewing a configured catalog before mapping or after price changes. Not for checkout, webhook or billing integration implementation, price-setting decisions or executing store changes.

Verified onCodex

Choose how each thing a business sells gets paid - a merchant of record, a payment gateway, or an invoice - by what is sold, who buys it, where, and in which currency, and check each choice against the provider's prohibited categories and the tax duty it leaves with the seller. Use when the user asks which payment provider to use, whether a product can be sold through a given provider, or how to split products across providers. Not for implementing one provider's API.

Turn a pricing idea into a testable hypothesis - who pays, for what unit of value, at what price and against which alternative, what result would change the decision, and the cheapest test - before anyone builds a pricing page. Use when the user proposes a price, a plan structure, a discount or a free tier, or asks how much to charge. Not for computing unit economics from real cost data or implementing billing.

Verified onClaude CodeCodexGrok CLI

Handle a refund, a partial refund or a payment dispute safely - read the current payment and refund state first, act once with an idempotency key, and tell the customer exactly once. Use when the user asks to refund, cancel, re-send a refund, answer a chargeback or fix a refund that may have failed. Not for designing a refund policy or writing payment integration code from scratch.

Implement or review a Toss Payments integration - payment confirmation, webhooks, virtual accounts, cancels and recurring billing - and cross-check every API detail against the official Toss Payments docs instead of memory. Use when the user builds, debugs or reviews checkout, payment confirmation, refunds or subscription billing with Toss Payments. Not for other payment gateways.

Verified onClaude Code

Turn a domestic price list (often tax-inclusive, in local currency) into a USD price list for a merchant of record or another tax-exclusive checkout - convert from the pre-tax price, hold the contribution margin under the provider's fees, set a floor where the fixed fee dominates, restructure prices the checkout cannot express, and keep annual plans consistent. Use when the user adds USD or international prices to an existing local price list. Not for setting prices from scratch.

Build a customer discovery kit that tests named hypotheses instead of collecting compliments - an interview questionnaire in blocks (situation, tool mix, procedures and authors, operating pain, trust and security, willingness to pay, close) with each question tagged by the hypothesis it tests, a record table with strong, weak and none defined, a decision rule written before the first interview, and a 15-minute demo script that marks what exists today versus what is still to be built. Use when preparing customer interviews, validating a B2B offer before building it, or turning hypotheses into interview questions and a go or no-go rule. Not for synthesizing feedback already collected or for writing a survey to existing users.

Write a one-page decision memo a busy decision-maker can act on - the decision needed and by when, the recommendation first, three real options compared on the same criteria with their cost of reversal, the evidence and what is assumed, the risks, and the exact approval being asked for. Use when a choice must go to an owner, founder, executive or operator for approval. Not for status updates or long reports.

Check a workflow or process flow, given as a diagram description, list of steps or arrows, for missing branches, dead ends and loops with no exit - enumerate every node and edge first, check that each decision has an edge for every outcome and each path reaches an end, list each loop with its exit condition and bound, and report findings by node name in a table. Use when someone pastes a flow, process, state or approval diagram and asks if anything is missing, whether it is complete, or to check the paths, exceptions and loops. Not for drawing a new diagram, for a sitemap, or for checking that the flow matches how the system actually runs.

Spars on a rough product or feature idea instead of listing ideas - asks the few questions that would change the answer, argues against each option, and narrows to one thing to test first. Use when the user asks to brainstorm, spar on, pressure-test or "think through" an idea, or asks what to try first. Not for framing a problem statement, ranking a given backlog, synthesizing feedback, or writing a finished spec, PRD or acceptance criteria.

Size one workflow as S, M or L before quoting a pilot that turns it into an AI skill - mark six criteria (where the rules live, number of rules and exceptions, evaluation cases, data and tool links, approval steps, security conditions), apply the count rule, show each mark with its evidence, and flag an L that will not fit a two-week pilot so it can be split before the quote. Use when preparing a quote or pilot proposal for building a skill or agent workflow, after a scoping call, or when someone asks how big a skill build is. Not for setting the prices themselves or for estimating a software project in hours.

Write a product requirements document in Korean - problem, goal and metric, users, scope and non-goals, requirements with testable acceptance criteria, screen copy and open questions - with every assumption marked. Use when the user asks for a PRD, a planning document or a requirements specification in Korean, or for acceptance criteria written in Korean. Not for brainstorming options or framing a problem that is still unclear.

Ranks a backlog with an explicit scoring method (RICE unless the user names another), showing each item's arithmetic, labelling every estimated input, flagging ranks too close to trust, and naming the one assumption that would change the top pick. Use when the user asks to prioritize, rank or score a list of features, bets or backlog items. Not for explaining a framework or generating new ideas.

Frames the problem behind a feature request or a vague goal - who is stuck, where, what evidence exists, how success would be observed, and what is out of scope - as a draft with its assumptions marked. Use when the user asks to frame a problem, write a problem statement, or work out what problem a requested feature solves before building it. Not for brainstorming options or writing a finished spec.

Draft and check the paperwork of a small project plan with the arithmetic shown - a charter (objective, in and out of scope, deliverables, assumptions, approvers), a work breakdown with an owner and estimate basis per task and a check that no deliverable lacks a task, a schedule check (critical path, buffer against the deadline, dependency cycles), and a change order that sets baseline against new scope, cost and schedule with an approval line. Use when someone asks for a project charter, a WBS or task breakdown, a check of schedule slack, or a change order or change request write-up, and gives the tasks, durations and deadline. Not for sizing a skill-build pilot (pilot-scope-sizing) or turning a feature brief into user stories (requirement-to-testable-stories).

Roll up a weekly status, daily brief, completion history or epic history from several sources - chat channels, tickets, repositories, spreadsheets - stating the time window and the sources queried or unavailable, citing a link and date for every item, separating done, in progress, blocked and slipped with the change since the last report, and flagging conflicts between systems instead of choosing one. Use when asked for a weekly status, a project brief, what was completed, how an epic changed, or a report built from several sheets or systems. Not for a coding-session progress update or a postmortem.

Break a feature request or technical brief into user stories whose acceptance criteria are checkable Given/When/Then statements, with vague words such as quickly, secure or easy flagged as unmeasurable and turned into a question for a value, dependencies and build order between stories, assumptions listed, and feasibility risks to verify with engineering. Use when someone pastes a brief, request or one-line requirement and asks for stories, tickets, acceptance criteria or a breakdown into work items. Not for a full Korean PRD (prd-writing-ko), for tracing stakeholder notes to sources (stakeholder-notes-to-requirements-trace), or for ranking the stories.

Consolidate scattered stakeholder notes, meeting transcripts and legacy solution-centric documents into business requirements that each carry a source quote, speaker or document and date, with business needs separated from proposed solutions, conflicts and duplicates across sources recorded, and gaps left as open questions instead of filled. Use when someone pastes meeting notes, transcripts or an old spec and asks for requirements, a requirements list, a process extraction or a traceable summary of what stakeholders said. Not for writing a PRD, interviewing customers, or prioritizing a backlog.

Synthesizes a batch of raw user feedback, tickets, reviews or interview notes into themes that cite the item ids behind them, ranks severity apart from frequency, escalates single severe reports, and separates what users asked for from the need behind it. For free-text survey answers it also accounts for every response, keeps outliers apart from themes, and turns each theme into an action with an owner role and a proposed date. Use when the user pastes several feedback items or survey comments and asks what users are saying, for themes, a summary for a product review, or an action plan from a survey. Not for replying to one user.

Review a new paid offer on paper against the organisation's own billing source-of-truth documents - map each product to an allowed model (subscription or one-time, feature grants and seat caps, stored value allowed or not, currency and tax rules, new products as catalog rows only, seller registration), list the decisions only the operator can make with a recommendation each, record approvals and changes as they come, and stop before catalog rows, provider variants, checkout or seller registration. Use when a new product, plan or paid offer is proposed, before pricing work turns into billing changes, or when asked to check an offer against the billing rules. Not for setting prices, implementing billing, or modelling entitlements in code.

Verified onClaude Code

Walk a product the way its users will, for real - define personas from beginner to expert, use a free and a paid test account, and run the public site, install or sign-up, login and the core feature end to end in a real browser or client, recording every point of friction as a complaint with evidence and separating launch blockers from polish. Use before a public launch or beta, when asked how complete the product feels, or for a consumer or user journey review. Not for a checklist of legal, payment and rollback readiness, and not for a code or design-token review.

Verified onClaude Code

Set a price for an offer that has none yet - a subscription tier, a maintenance or care retainer, an add-on, or an advertising rate card - from three inputs read separately (what comparable offers charge for the same job, the cost per unit with the provider's fees, and the features only this offer has), then charge the market rate for parity and a priced premium for the difference, with a margin floor and a re-pricing trigger. Use when the user asks what to charge for something new, or whether a proposed price holds up. Not for testing an existing price (use a pricing hypothesis) or for converting a price list into another currency.

Verified onCodex

Model what each plan entitles a customer to - features as boolean grants, limits as metered allowances with their unit, reset period and overage behaviour, trials and grandfathered customers - as one table the product checks at runtime instead of plan names scattered in code, and walk the upgrade, downgrade and cancellation cases. Use when designing or changing pricing plans, adding a paid feature, or when plan checks are spread through the codebase. Not for setting the prices themselves.

Verified onClaude Code

Settle a design or product decision with a panel of read-only expert seats - write a measured evidence packet first, run 3 to 6 seats (risk, governance, cost, measurement, security, a consumer persona) plus one mandatory dissent seat, each citing only file:line or measured evidence, merge them into one recommendation with options, ask the operator for one confirmation in plain words with cost flagged, then record the result as a decision record or plan. Use when the user asks for an expert panel, an expert review or meeting, a recommendation on a contested design choice, or when several viewpoints conflict. Not for a quick factual question or for reviewing a code diff.

Verified onClaude Code

Plan a launch in the Korean market across the channels people there actually use - Naver search, blog and place listings, KakaoTalk channel, Korean communities and cafes, app and extension stores - with a message per audience, sequenced pre-launch, launch and follow-up steps, owned assets ready, community rules respected, and success measured per channel. Use when launching a product, feature or brand to Korean users. Not for paid ad buying.

Verified onClaude Code

Check a product or feature is ready to launch across the areas that fail quietly - legal and policy text, payments and refunds, status page and monitoring, support readiness, data and privacy, and a tested rollback - by running each area as an independent check (in parallel subagents where available), each returning evidence and a go, conditional or block. Use before a launch, a paid feature going live, or a public announcement. Not for code review.

Verified onClaude Code

Write the customer-facing offer for a paid pilot that turns one workflow into an AI skill - what the pilot is and includes, the size and the price exactly as the company supplied them, the timeline from consultation to handover, what the customer prepares, the before and after measures, payment, a link to the security and data-flow disclosure, and the options after the pilot - leaving any commercial term the company has not set as "as stated in the quote" instead of inventing it. Use when drafting a pilot quote, proposal or offer letter for a customer, or turning an approved pilot scope into a document the customer signs off. Not for sizing the workflow, setting prices, or writing the contract itself.

Verified onClaude Code

Review a portfolio of products or brands to decide where to invest, what to maintain and what to wind down - each product scored on the same measured inputs (usage trend, revenue, cost to run, maintenance load, risk, strategic fit), unknowns labelled, the decision per product stated with what would change it, and the wind-down cost counted. Use when someone runs several products, brands or side projects and must choose where time and money go. Not for prioritising features inside one product.

Verified onClaude Code

Derive a per-unit cost line (per character, second, image or call) from an AI or API provider's published prices - the official price page with URL and date, conversions measured with a real tokenizer or labelled as estimates, a formula in one fixed small currency unit, estimated prices kept on their own board graded estimate, cost rows that inherit the board's grade, and changes as forward-only migrations the owner approves. Use when adding a provider or model to a cost ledger, pricing a feature billed in a different unit than the provider charges, or when a cost row's evidence is questioned. Not for setting customer prices or modelling margins.

Verified onClaude Code

Run a roadmap as a sequence of gated bets - each item framed as a problem with evidence, sized and ranked with its fragile assumption named, given an exit criterion and a kill criterion, and only started when the previous gate's evidence is in - and record each go, hold or kill with its reason. Use when planning a quarter or a product roadmap, deciding what to start next, or reviewing whether an in-flight item should continue. Not for a single feature's PRD.

Verified onClaude Code

Model unit economics for a product with AI costs - revenue per unit, variable costs per unit (model tokens per task at measured volumes, payment fees, hosting, support), contribution margin, acquisition cost and payback - with every input sourced or labelled as an estimate and the margin's sensitivity to the inputs that move it most. Use when pricing an AI feature, checking whether a plan makes money, or deciding usage limits. Not for company-level financial statements.

Verified onCodex

Turn pasted excerpts of public pages (competitor sites, a prospect's pages, pricing or positioning copy) into a comparison brief for sales or marketing - the same dimension rows for every company, each claim tied to its source page and date, stated facts kept apart from inference, and a gap list of what the excerpts do not say. Use when someone pastes page text and asks for a competitor comparison, a positioning summary or an account brief. Not for browsing the web, for facts that are not in the pasted text, or for judging a person.

Verified onClaude CodeGrok CLI

Mine a user's sent-email history for external contacts, last topics and promised follow-ups, resolving relative dates against each email's own date, checking the later thread for a completing message, and marking each commitment done, open or unclear with the quoted line. Use when someone wants forgotten contacts, relationship topics or never-completed follow-ups recovered from past email, or an end-of-day open-loops list. Not for sending messages, scoring leads, or analysing internal or automated mail.

Verified onClaude CodeGrok CLI

Turn a sales discovery or demo call transcript into CRM framework fields (such as MEDDICC) backed by short transcript quotes, a list of open gaps as next-call questions, and a follow-up email built only from commitments actually stated. Use when a rep has a call transcript and must update the CRM and write the follow-up. Not for forecasting, deal scoring, or inventing budget, authority or timeline the customer did not state.

Verified onClaude CodeGrok CLI

Write an architecture decision record (ADR) so a later reader can tell whether it still applies, including what was rejected and what would reverse it. Use when the user asks for an ADR or a decision record, or when a choice closes off alternatives or is expensive to undo. Not for plans still under discussion.

Verified onClaude Code

Recover a plan document that a read-only planning subagent returned only inside its final report, verbatim from the session transcript JSONL, save it under the project's plans directory with its naming convention, and diff it against the report to prove nothing was paraphrased. Use when a subagent or plan agent produced a full plan but could not write files and the plan must be saved. Not for writing a new plan from scratch.

Verified onClaude Code

Write a plan document that states what is proposed, what was measured, and what is still unverified, and that can be superseded without being deleted. Use when starting multi-step work or recording an investigation.

Write a source-of-truth, rule or convention document that can be proven stale, by pairing every rule with a command that checks it. Use when the user asks to document a rule, an invariant or a coding convention, or to write a source-of-truth document that later work must not quietly contradict.

Verified onClaude Code

Bring an existing pile of design documents into a schema without losing content or inventing values. Use when backfilling frontmatter, normalising status vocabularies, splitting an over-long document, or reviewing a stale one.

Verified onClaude Code

Work through a dossier of existing documents and decide each one's status, relationships and placement by reading it. Use when normalising an accumulated document set into a schema.

Verified onClaude Code

Triage a folder of plans against the current code - read each plan and the tree, judge it applied, partial, obsolete or active with file and commit evidence, then archive what is done or dead, split the unfinished remainder of partial plans into a small active list, and promote lasting rules and decisions into source-of-truth documents or decision records. Use when a plans directory has grown stale, when nobody knows which plans are still live, or before a planning round. Not for writing a new plan or for normalising document filenames and layout.

Verified onClaude Code

Bring design documents (sources of truth, decision records, plans, specs) up to date with recent code changes by reading - group the commits since the last sync, read each diff and each candidate document with judgement rather than matching strings, decide per document whether its statements are now wrong, incomplete or still true, and edit only those, citing the commits. Use when the specs have drifted after a run of changes, when the user asks to sync or refresh the docs or SoT after recent work, or on a regular cadence. Not for migrating a whole pile into a new layout or for writing a new document from scratch.

Where design documents live, what they are named, and which of those rules can be changed. Use when adding a document kind, placing a new document, or deciding whether a naming rule is worth a migration.

Verified onClaude Code

Draft an information architecture sitemap from a project description by deriving every node from a stated user task, limiting depth to a stated maximum with each level labelled, emitting text-based diagram source (Mermaid) that can be checked, listing which tasks and features map to which nodes, and ending with the open questions the description could not answer. Use when someone gives a project or feature list and wants a first sitemap or IA diagram. Not for visual design or wireframes, for validating a structure with users, or for writing content.

Verified onClaude Code

Index long interview transcripts by session, theme and timestamp, and pull quotes for a report that are character-for-character in the source, then audit every quote against the transcript. Use when summarising or theming user-interview transcripts, finding where a topic was discussed, or collecting quotes for a deck or findings report. Not for writing interview guides or analysing survey data.

Verified onClaude Code

Calculate System Usability Scale scores from raw questionnaire responses by checking each response sheet is complete and in range, applying the scoring rule the user supplies item by item, showing the per-respondent arithmetic, and reporting the mean, median and number of respondents scored with the excluded sheets named. Use when someone pastes SUS answer sheets and wants scores, or asks to recheck a SUS calculation. Not for choosing or writing the questionnaire, for interpreting what a score means for a product, or for other survey scales.

Verified onClaude Code

Read photos, scans, multi-column PDFs, charts, signs, device screens and clothing images for a blind or low-vision user - answer the question asked first, recover reading order, state chart axes, units and trend, separate what is seen from what is guessed, and guide device menus one confirmed step at a time. Use when a blind or low-vision user shares an image or document and needs the text, order, trend or color rather than a scene description. Not for producing alt text for publishing or for image generation.

Verified onClaude Code

Review or draft image alternatives for a web page, document or publication with the caption, surrounding text, link target and audience in view - classify each image as decorative, informative, functional, text-bearing or complex, write the alternative that purpose needs, drop invented emotions, identities and chart readings, and check the alternative that the exported file actually carries. Use when an assistant, CMS or tool produced alt text drafts, or alt text must be written for images whose page context is available. Not for describing an image to a blind user, for navigation help, or for certifying accessibility compliance.

Verified onCodexGemini CLIGrok CLI

Write task routes for screen reader and keyboard users and shape long answers for heading navigation - name controls by label with real key commands for the stated reader and platform, mark each key command or control name as confirmed or unverified, end every step with a non-visual completion signal, and give long replies stable numbered headings with a one-line summary first. Use when a screen reader user asks how to do a task in an app, site or device, or wants a long answer they can jump through by headings and re-read. Not for reading images or documents aloud, for auditing a site's accessibility, or for writing alt text.

Compare an exported PDF's intended sequence, tag tree, form tab order and reader observations by element ID. Use when checking PDF reading order, tagged forms or accessibility export reports. Not for judging structure from screenshots or certifying whole-document accessibility.

Verified onCodex

Draft an internal audit program or per-control test steps for a process - start from risks and control objectives, map each control to attribute, population, sample basis and evidence, write steps another auditor can execute (who, which document, pass or fail criterion), list the data requested from the auditee, and mark sampling and risk ratings as the auditor's judgment. Use when an internal auditor asks for an audit program, test procedures for a control, or a data request list for a process. Not for audit opinions, concluding that a control works, setting final sample sizes as fact, or citing standards not supplied.

Verified onClaude Code

Turn a process walkthrough transcript into a control narrative and write findings from it - per step the owner, backup, system, frequency, evidence and approval threshold with the transcript line quoted, gaps listed where the transcript is silent instead of filled in, and each finding written as condition, criteria, cause, effect and recommendation with an evidence reference, using the criteria text the user supplies. Use when an internal auditor has walkthrough notes or a transcript and needs a control narrative, or has verified issues and needs executive-level finding text. Not for audit opinions, concluding a control works or fails, inventing frequencies, owners or criteria, or citing standards the user did not supply.

Verified onClaude Code

Review a completed audit workpaper, possibly AI-assisted, against the evidence and procedure records behind it - trace each factual statement and test result to a source location, separate management representations from inspected evidence, reconcile population, sample and exception counts to the attachments, flag procedures described as done with no execution record, challenge the proposed conclusion against contradicting evidence and scope limits, and return reviewer questions with an evidence-to-conclusion matrix. Use when an audit lead or reviewer must check a finished workpaper before sign-off. Not for designing the audit program (audit-program-test-design), drafting walkthrough narratives (audit-walkthrough-and-finding-writeup), issuing an opinion, or deciding evidence sufficiency; the responsible auditor decides.

Verified onCodexGemini CLIGrok CLIAntigravity

Explain a month-over-month cash flow movement or a spiking expense category with a bridge that sums - start value to end value by driver, any residual named as unexplained, the spike drilled by vendor and transaction into one-off, timing and price or volume, the filter shown so numbers can be re-run, and each cause marked evidenced or hypothesis. Use when asked why operating cash fell or why one expense category jumped between two months. Not for forecasting, investment or accounting conclusions, or a ledger rollforward.

Verified onClaude Code

First-pass review of a borrower's financial statements for a credit analyst - compute leverage, coverage, liquidity and margin ratios with the formula and period label shown, check that statements foot and that equity rolls forward, flag restatements, related-party items and unusual notes, and separate facts from the analyst's judgments in a memo section. Use when an analyst or credit officer pastes borrower figures and wants a first review or help refining a credit memo. Not for assigning a rating, recommending approval or decline, pricing, or covenant decisions.

Verified onClaude Code

Turn pasted bank statement text or a pile of uploaded client documents into structured rows and a file inventory with a tie-out - keep the original text, check opening plus the sum of rows against closing per statement, classify each file by type, period and entity with the header text as evidence, and mark unreadable digits instead of guessing. Use when a bookkeeper or tax preparer receives statements or scans with meaningless names and needs rows and a rename plan. Not for categorizing transactions into accounts, tax advice or filing, or a month-end rollforward.

Verified onClaude Code

Roll a monthly reconciliation, ledger pivot or intercompany balance match forward to a new period without stale dates or broken balance links - list every period-dependent date, range and filter, tie opening to prior closing, match counterparty pairs with currency and sign, list unmatched items, and keep a change log for the reviewer. Use when a preparer copies last month's reconciliation file for the new close. Not for posting journal entries, approving a reconciliation, or deciding how an item is accounted for.

Verified onClaude Code

Build a sourced company dossier for an analyst from filings and earnings-call transcripts - every fact tied to a named document, section and date, segment and geography exposure tabulated with fiscal-year labels and units, management statements kept as claims apart from reported figures, and a list of what was not found or not opened. Use when an analyst asks for a management profile, a segment or geography exposure table, or a business and risk summary of a listed company from supplied annual reports or call transcripts. Not for investment recommendations, valuation, price targets, or confirming that a figure is correct.

Verified onClaude Code

Compare QIF exporter date/year and field-specific numeric conventions with independent controls and actual stored importer rows. Use when totals match but imported transaction dates or amounts may be misinterpreted. Not for financial advice, postings, bank duplicate reconciliation or guessing ambiguous locale/currency.

Verified onCodex

Research a narrow tax question into a map of primary authority with what was and was not opened, an amendment check with its date, and a client memo split into answer, authority, assumptions and open questions, with payment steps taken only from the official source. Use when a tax preparer or advisor asks whether a provision applies, wants an assistant's cited section checked, tracks a legislative change, or needs a plain-language payment explanation. Not for tax advice, filing or payment instructions, deciding that a deduction applies, or replacing review by a qualified tax professional.

Verified onClaude Code

Check a three-statement financial model's integrity before trusting its projections - confirm the opening balance sheet balances, independently recompute net income, retained earnings and ending cash from the supplied inputs, list differences by period, trace an imbalance to the input row that created it, and report what was and was not recomputed. Use when a modeler has an AI-built or inherited model and asks whether the balance sheet balances and cash ties. Not for forecasting, valuation, choosing assumptions, or investment conclusions.

Verified onClaude Code

Assemble a denied-claim appeal packet for a clinic from the denial letter and the chart - every fact quoted with its source and date, the original denial reference and the resubmitted claim number kept separate and conflicts flagged, each procedure tied to the recorded finding that supports it, missing documentation listed, and no clinical fact added. Use when a dental or medical office has a denied or questioned claim and wants an appeal packet or necessity narrative drafted. Not for clinical judgement, deciding the appeal will succeed, writing coding advice, or identifiable patient data.

Verified onClaude Code

Reorganise a clinician's own brief session notes or stated goals into a required progress-note or treatment-plan template (SOAP, DAP, goals and interventions) by placing each fragment in a field, marking empty fields NOT DOCUMENTED, and ending with a list of every phrase that is the model's wording or inference. Use when a clinician or practice admin pastes short notes plus a template and wants a draft record. Not for diagnosis, risk assessment, billing codes, treatment advice or summarising a record for someone else.

Verified onClaude CodeCodex

Build a prior authorization evidence packet from a de-identified treatment history and the payer's stated criteria - a criterion-by-criterion table with dose, dates, outcome and source, MISSING marked wherever the record is silent, denial reasons sorted into not submitted, submitted but not cited, and genuine non-coverage, and renewals carried forward with what changed. Use when staff or a clinician prepare, appeal or renew an insurance prior authorization. Not for clinical judgement, choosing a treatment, writing a medical necessity conclusion, or anything with identifiable patient data.

Verified onClaude Code

Build Boolean search strings for candidate sourcing from the job requirements the user states - the platform's operators, grouped synonyms and title variants, exclusions, a narrow and a broad variant, and a note on platform limits - while leaving out protected characteristics and their proxies. Use when someone needs a search string for LinkedIn or a similar people-search platform from a job description or hiring-manager requirements. Not for ranking, scoring or screening out candidates, profiling a named person, or deciding who to contact.

Verified onClaude Code

Compare candidate office locations from aggregated employee home areas - work only on area-level counts, apply a minimum group size and fold small groups into an "other" row, weight each area by its count and its travel time (not straight-line distance) to each candidate, report the weighted mean and the share above a stated commute limit, and list assumptions and what the analysis cannot show. Use when an HR, workplace or facilities planner asks which office site suits the home distribution of staff. Not for working with individual home addresses, naming or locating any employee, deciding a relocation, or advice on employment or legal consequences.

Verified onClaude Code

Lay out interview notes, intake transcripts or a long complaint for a human investigator as a statement-by-statement table with speaker, exact quote and location, then separate allegations, points several statements agree on, contradictions and unknowns, keeping names and dates exactly as given and marking missing ones. Use when someone asks to summarise, compare or find themes across witness statements, interview notes or a written complaint. Not for judging credibility, deciding whether a policy was broken, recommending discipline, or legal advice, which stay with the investigator and a qualified HR or legal reviewer.

Verified onClaude Code

Review an internal policy or procedure draft against the user's own checklist or source rule - required elements listed first, each checked in the draft with the clause quoted or marked missing, contradictions between clauses and with the source flagged, site-specific facts left as blanks for the owner, and what could not be verified stated. Use when someone has a draft policy (absence, expense, safety, AI-use or similar) and a rule or checklist it must satisfy. Not for writing the policy's content, giving legal advice, or approving the policy.

Verified onClaude Code

Split long or overnight shifts into clock intervals and count paid, overtime and premium hours under the rule the user pastes - unpaid breaks removed, each interval tagged once, overlaps shown explicitly, and the rule restated as applied. Use when someone has a shift schedule that crosses midnight or runs long and wants overtime and premium (night, weekend) hours calculated from a stated rule. Not for deciding what the law requires, setting pay rates or computing wages beyond hours, or approving payroll.

Verified onClaude Code

Compute aggregate workforce metrics such as 90-day attrition, turnover and labor-hours or payroll pivots from HR rows the user pastes - numerator, denominator, cohort and period defined first, terminations, transfers and rehires handled explicitly, a minimum group size applied, and row counts reconciled to the source. Use when someone needs attrition, turnover or hours and payroll totals by department or group from HR data. Not for ranking, scoring or judging individual employees, predicting who will leave, or defining a product metric (see metric-definition in data-analytics).

Verified onClaude Code

Check repair-estimate rows (room, trade, item code, quantity, unit) - roll quantities up by trade and unit per room and in total with a source row for every figure, never adding mixed units, match an original and a supplement estimate by item code into added, removed and quantity-changed, recompute stated measurements and subtotals, and flag rows that conflict with quoted guideline text. Use when an adjuster or contractor pastes estimate rows and wants quantity totals, a supplement comparison or a pre-submission check. Not for pricing, coverage or payment decisions, approving a supplement, or fraud conclusions.

Verified onClaude Code

Compare emergency-mitigation and restoration estimate lines by location, period and scope to mark possible overlaps for the handler to confirm, and match living-expense receipts to a claimed line by amount, date range and item with recomputed totals, separating matched, unmatched, duplicate-looking and out-of-period items with the document reference of each. Use when a claims handler or adjuster pastes mitigation and restoration estimates or receipts and a claim list and wants possible double billing or unmatched receipts found. Not for coverage or payment decisions, fraud conclusions, or approving a claim.

Verified onClaude Code

Review a commercial client's actual operations against the class on file, the underwriting guide and the policy summary - list each operation and exposure, compare with the class and guide text with their dates, list the exclusions and endorsements that bear on each exposure, and mark every gap as a question for the licensed professional. Use when a broker, agent or account manager asks whether a business's activities fit its class code and coverage, or what coverage gaps to check. Not for deciding coverage, recommending a carrier or limits, or telling the client they are covered or uncovered.

Verified onClaude Code

Normalize several carrier proposals or benefit summaries into one comparison table - map each carrier's coverage, limit, deductible or retention and exclusions to common columns, keep each term's own basis (per occurrence, aggregate, per claim) and unit, mark every term a proposal does not state as not stated, and quote the source page for each cell. Use when a broker or buyer pastes two or more proposals and asks for a side-by-side comparison. Not for recommending a carrier, judging adequacy of cover, pricing, or coverage and legal conclusions.

Verified onClaude Code

Review ad copy before it runs - every factual claim backed by evidence you can show, no superlative or comparison you cannot prove, sponsorship disclosed, category rules for food, cosmetics and health products, and one clear promise per ad - then return a marked-up revision with the reason for each change. Use when the user asks to check, review or fix an ad, a promo banner, a sponsored post or a product claim before publishing. Not for writing an ad from scratch or for in-app UI copy.

Decide how AI-generated or AI-assisted content must be labelled where it is published - platform rules (YouTube, Instagram, Naver, app stores), Korean rules on AI-generated content and advertising, and sponsorship disclosure - and write the label, the metadata and the record that proves it. Use when publishing images, video, voice or text made with generative AI, or when a platform asks whether content is synthetic. Not for detecting whether content is AI-made.

Verified onClaude CodeCodexGrok CLI

Judge the legal risk of collecting data from a website or app in Korea - unauthorised access, database rights, the data and free-riding clauses of the Unfair Competition Prevention Act, personal data in what is collected, and contract terms - with current article text and court decisions retrieved, not remembered. Use when the user plans or reviews scraping, crawling or bulk collection of another service's data for a Korean product. Not a substitute for counsel in an actual dispute.

Verified onCodex

Review an online sales or subscription flow against Korea's Act on Consumer Protection in Electronic Commerce - seller information, pre-purchase disclosure, cancellation and withdrawal rights and their exceptions, automatic renewal and free-trial conversion notices, and refund timing - citing retrieved article text. Use when the user sells goods, digital content or subscriptions to Korean consumers and asks whether the checkout, refund policy or renewal flow is compliant. Not for business-to-business contracts.

Verified onCodexGrok CLI

Review a feature, form, vendor setup or data flow against Korea's Personal Information Protection Act - collection and consent notices, entrustment to processors, provision to third parties, overseas transfer, minors and sensitive data - citing the current article text. Use when the user asks whether a Korean product's handling of personal data is lawful, or needs a consent notice, privacy policy section or vendor review for Korean users. Not a substitute for counsel on a dispute.

Verified onCodex

Draft terms of service and a privacy policy for a Korean online service from how the service actually works - the data it handles, its processors and overseas transfers, payments and refunds, and user content - with every legal requirement cited from retrieved text and every product fact confirmed. Use when the user needs terms of service or a privacy policy drafted or updated for a Korean service, in Korean. Not for reviewing an existing data flow (use pipa-review) or for contracts between companies.

Verified onCodex

Check the rights behind a voice before it is used in a product or ad - a voice actor's contract scope, consent and limits for a cloned or synthetic voice, the TTS provider's licence for the use, the publicity-right protection of a person's voice under Korean unfair competition law, and AI disclosure - and list what is missing. Use when using a human voice recording, a voice clone or a TTS voice commercially, or when someone asks to make a voice that sounds like a specific person. Not for audio quality.

Verified onCodex

Review a contract draft, including a model-generated one, against an approved precedent clause by clause, test the formation elements and the boilerplate list, flag defined terms that are used but undefined, and turn each reviewer instruction into two or three redline options with the other clauses each touches. Use when someone asks whether a draft is complete, what is missing versus the precedent form, or how to implement an instruction such as making a clause mutual or capping liability. Not for legal advice, deciding validity or enforceability, approving or signing a draft, or negotiating terms.

Verified onClaude Code

Prepare a pre-signature brief of a contract for a non-lawyer signer - a table of term, payment, termination, renewal, liability cap, indemnity and governing law with the clause number and the quoted text beside every line, the clauses that look unusual or one-sided, standard clauses not found in the supplied text, the dates a notice period creates, and questions to put to counsel. Use when an executive, founder or assistant wants the highlights of an agreement before it is signed. Not for legal advice, saying whether to sign, validity or enforceability, negotiating, or comparing a draft with a precedent (use contract-draft-completeness-review).

Verified onClaude Code

Compare a witness's deposition testimony with written answers or other statements by aligning them by topic into a matrix with a locator on each side, classifying each pair as consistent, apparent conflict or different scope with both quotes, listing the context that could explain each apparent conflict, and listing statements with no counterpart. Use when an attorney or paralegal needs candidate inconsistencies between approved, de-identified statements for review. Not for judging credibility, calling anyone untruthful, deciding significance, impeachment strategy or legal advice.

Verified onClaude Code

Check draft discovery requests against a complaint by extracting each claim and its elements, building a requests-by-element coverage matrix, listing elements no request reaches and requests tied to no pleaded fact, flagging boilerplate, and proposing candidate objections per request with the trigger words and the rule to confirm. Use when an attorney or paralegal reviews draft requests for production or interrogatories before service, or prepares candidate objections to requests received. Not for legal advice, citing rules or cases from memory, or deciding what to serve or object to.

Verified onClaude Code

Compare a revised contract, policy, renewal or redline against the previous version by aligning clauses, listing every change with old and new text, catching removed, moved and changed-number text, rating the impact per change and who bears it, and marking items for legal or broker review. Use when someone asks what changed in a renewal, amendment or new draft, or whether a revision is safe to accept from its summary. Not for legal or insurance advice, approving the revision, or negotiating terms.

Verified onClaude Code

Convert a day's work notes into draft billing time entries that follow the client billing guidelines the user pastes (no block billing, minimum increments, banned words, required wording), keeping the stated total, citing which rule drives each change and listing the entries the guidelines would reject. Use when someone pastes work notes with hours plus the client's billing rules and wants entries ready for review. Not for legal advice, deciding what is billable, fee disputes, or reviewing OCR or transcript text.

Verified onClaude Code

Build an issue-scoped medical chronology, a medication table and a records-completeness list from de-identified records for a legal matter - issue and date window stated first, every row with a source locator and verbatim snippet, fills kept apart from prescriptions, ambiguous entries flagged not resolved, and a missing-items table by provider, period and document type. Use when a paralegal, records reviewer or attorney needs a chronology or a check of which provider records and bills are still outstanding. Not for medical opinions, causation or damages conclusions, or identifiable patient data.

Verified onClaude Code

Review OCR text of scanned records or an automatic transcript of poor-quality audio and produce a ranked list of risky spans (numbers, dates, doses, names, negations, speaker turns) with locator, original span and reason, without rewriting the text. Use when someone must rely on OCR or speech-to-text output where a single wrong digit, name or speaker would matter. Not for correcting the text, summarising it, or deciding what a record means.

Verified onClaude Code

Triage Korean customer support tickets - classify by type and urgency, pull out the facts needed to act (account, order, error, time), detect duplicates and incidents across tickets, route to the right owner, and draft a polite reply in haeyo-che that promises only what the product does. Use when sorting a batch of Korean support messages or answering one. Not for writing help-center articles.

Verified onClaude Code

Run a blameless incident postmortem end to end - a UTC timeline built from logs and messages, impact in numbers, the root cause and contributing factors traced past "human error", and action items with owners, dates and a way to verify each - and publish it. Use when the user asks for a postmortem, an incident review or an RCA after an outage or a near miss, in any language. Not for live incident response while the outage is still ongoing.

Verified onCodex

Write a read-only operator runbook that someone on call can follow - a trigger, the checks in order with the exact command or dashboard and what a healthy result looks like, the decision each result leads to, when to escalate and to whom, and nothing that changes state without an explicit, separate step. Use when documenting how to check a service, a queue, a payment flow or an admin console, or when turning tribal knowledge into a checklist. Not for incident communication.

Verified onClaude Code

Answer a customer question or take a confirmation from knowledge-base and reservation data without ignoring current deadlines, temporary notices or handoff conditions - cite the source, let an unexpired notice override the standing answer, escalate exceptions, and read numbers back digit by digit. Use when replying to guest or customer messages for a hotel front desk, booking or extension request, support chat or voice assistant that must follow a stated policy and deadline. Not for sorting a batch of tickets or writing help-center articles.

Verified onClaude Code

Write status page and customer updates during an incident - what is affected in user terms, what users can do now, when the next update comes, no guessed causes or deadlines - and the closing note when it is resolved. Use when posting or drafting an incident update, a status page notice or a maintenance announcement. Not for the internal postmortem.

Verified onCodex

Write an admin procedure for a named product and exact version only from the vendor documentation supplied - record the product, version and doc version for every step, prefer release notes over older guides where commands changed, mark each step "documented" or "not in the supplied docs" instead of filling gaps with plausible commands, separate read-only checks from state changes, and list what the admin must confirm. Use when an admin asks for steps, commands or settings for a specific product release. Not for reviewing finished scripts or for incident runbooks (use operator-runbook).

Verified onCodex

Check a drafted claim set against its specification and drawing reference list - find terms introduced with the or said and no earlier a or an, group term variants across claims and description (for example fastener and attachment member), and build a reference-numeral table from the text and from the drawing list to show numerals present in only one of them, with every location quoted and the wording choice left to the drafter. Use when a drafter wants a consistency pass on claims, description and drawing numerals before filing or sending to the attorney. Not for patentability, claim strategy, claim scope advice, or adding new matter; the patent attorney or agent decides.

Verified onCodexGrok CLIAntigravity

For a patent practitioner, map each rejection ground in an office action to the claim elements and to the cited passages the user pasted (with paragraph or column locations), check every proposed argument or amendment against the specification text supplied and mark text with no support, relate candidate continuation claims to the existing claims and flag possible new matter, and end with what was not verified for the attorney or agent to decide. Use when someone pastes claims, an office action, cited references or specification passages and wants element-level mapping, a response outline or continuation candidates. Not for patentability, validity, infringement or freedom-to-operate opinions, for choosing prosecution strategy, or for any legal conclusion; the registered patent attorney or agent decides.

Verified onClaude Code

Compare a list of patent numbers, titles, assignees, inventors, dates and kind codes - produced by a model or extracted from documents - field by field against the source records the user pasted (register extract, publication front page), marking each field as exact match, partial match with the variant quoted, mismatch with both values quoted, or not in the sources, and never completing a missing field from memory. Use before such a list goes into a conflict check or a citation list. Not for searching for patents, deciding conflicts of interest or freedom to operate, or any patentability view; the patent attorney or agent decides.

Verified onCodex

Retire a redirect on a Cloudflare zone and answer 404 instead on the free plan - a small module Worker on the apex and www routes, attached before the redirect rule is deleted so there is no 522 gap, then verified with curl. Use when a domain should stop redirecting and return not found, when parking or retiring a Cloudflare zone, or when a WAF custom response is refused as not entitled. Not for writing a Worker's application logic or reviewing wrangler bindings.

Verified onClaude Code

Review a Cloudflare Workers project's bindings and configuration - service bindings instead of public URLs between workers, secrets as secrets not vars, environments that do not share production resources, compatibility date and flags, routes, observability and limits - before deploy. Use when reviewing or changing wrangler.toml or wrangler.jsonc, adding a binding, KV, D1, R2, queue or durable object, or debugging a worker that works locally but not deployed. Not for writing the worker's business logic.

Verified onCodex

Check a Chrome extension against the Chrome Web Store program policies before submission or after a rejection - single purpose, permissions justified one by one, user data disclosure and limited use, remote code, deceptive behaviour, and the listing itself - and map a rejection notice to the exact policy and fix. Use when preparing a Chrome Web Store submission, answering a review rejection or adding a permission to a Manifest V3 extension. Not for general extension debugging.

Verified onCodex

Set or tighten a Content-Security-Policy on a page behind a CDN that injects scripts, by loading the live page in a headless browser, collecting CSP violations and network requests, allowing only what the owner wants, hashing the site's own inline script from its source text, and re-loading to confirm the beacon succeeds and only deliberate blocks remain. Use when adding, tightening or debugging a Content-Security-Policy header, or when an analytics beacon or inline script is blocked by CSP. Not for general security reviews or CORS errors.

Decide whether apparently unused code or a legacy service should be kept or retired, checking dynamic callers and external consumers and separating shutdown from permanent data disposal. Use when an unused-code report prompts removal, when stopping a legacy process or schedule, or when deciding what a cleanup may delete. Not for refactoring code that is in use or for dependency upgrades.

Verified onClaude Code

Map a requested feature to existing implementation and actual wiring before planning a duplicate. Use when adding functionality, choosing reuse versus new code, or a missing UI or search result is taken as evidence that a capability does not exist. Not for validating a supplied code-review finding or redesigning a working feature without a requested gap.

Verified onCodex

Design or review an MCP server's tools - few tools with clear names and descriptions written for the model, input schemas that reject bad calls, stable ids instead of positions, read-only and destructive annotations, paging and size limits, errors the model can act on, and auth that keeps credentials out of the conversation. Use when building an MCP server, adding or renaming tools, or when models misuse a server's tools. Not for building an MCP client.

Verified onCodex

Review a Chrome Manifest V3 extension's architecture - a service worker that can be killed at any time, state that survives it, messaging between contexts, content scripts isolated from the page, permissions and host access kept minimal, and no remote code - before release. Use when building or reviewing a Manifest V3 extension, migrating from V2, or debugging an extension that stops working after idle. Not for store listing or policy review.

Verified onCodex

Review a PostgreSQL schema migration for production safety - which statements take locks that block reads or writes, how long they hold them, CONCURRENTLY and NOT VALID patterns, backfills in batches, lock and statement timeouts, and the deploy order between code and schema. Use before running a migration on a live Postgres database, or when a migration caused downtime. Not for ClickHouse or query tuning.

Verified onCodex

Move a public static site to a new host or generator without breaking a URL - snapshot the live site path for path, build every URL form as a real file, compare a local container against the snapshot for status, media type and bytes, check renderer parity when the Markdown engine changes, and cut over DNS only after it passes. Use when migrating a static site, docs or blog to a new host, server or static site generator, or replacing its Markdown renderer. Not for redesigning the site or setting up a new site from scratch.

Verified onClaude Code

Prepare draft customs entry lines from pasted commercial invoice rows with a row-by-row crosswalk to the source, recompute each line amount and the invoice total, compare quantity and unit with the source, leave unsupplied fields blank, and check scientific names on a plant-health form against a reference list the user supplies. Use when a broker or importer pastes invoices or a plant-health form and wants entry lines or a form check prepared. Not for classifying goods, deciding origin or duty, or filing; a licensed customs broker or the authority decides those.

Verified onClaude Code

Compare dedicated-carrier bids that mix a fixed weekly charge with a per-mile rate, or distribution-center lane scenarios, by converting every bid to weekly cost, cost per mile, per trip and per stop under one stated volume, counting empty return legs as billed miles, finding the break-even volume between bids and listing which fees each bid leaves out. Use when a buyer or logistics planner must choose between fixed-plus-variable bids or one-way and round-trip lane layouts. Not for contract negotiation, carrier qualification, or an award decision, which stay with the buyer.

Verified onClaude Code

Compare carrier or freight bids that mix fixed and per-mile pricing by normalising each to a total cost per lane at stated volumes, separating fixed, variable and accessorial lines, flagging missing terms, running a volume or pack-size sensitivity and recomputing every total before any recommendation. Use when a buyer is choosing between carrier bids or modelling container utilisation scenarios. Not for contract negotiation, carrier qualification, or a final award decision, which stay with the buyer.

Verified onClaude Code

Prepare a provisional freight class or tariff classification for an unusual shipment by naming the classification system and edition, collecting commodity description, dimensions, weight and packaging, computing density, applying the density table the user supplies, and stating that the official schedule entry and a licensed broker or carrier decide. Use when a shipper or buyer needs a candidate freight class or classification evidence for a shipment. Not for filing a classification, customs declarations, duty advice, or inventing item numbers.

Verified onClaude Code

Extract a product-specific preferential origin rule only from rule text the user pastes, quote each alternative verbatim, list the facts still needed to apply it (bill of materials with headings and values, processes, price basis, calculation method), show any arithmetic from supplied numbers with the formula marked as assumed, and state agreement version and date to be checked. Use when a trade user asks whether a product meets a preferential origin rule. Not for deciding origin, tariff classification, duty, or certificates; the authority and a qualified customs professional decide those.

Verified onClaude Code

Turn scattered negotiation notes or a draft RFP scope into a structured brief - per term a target, a walk-away and alternatives, every note labelled fact or guess, scope inclusions, exclusions and evaluation criteria listed, ambiguous scope flagged and open questions collected, with no figure invented. Use when a buyer pastes loose notes before a supplier negotiation or wants an RFP scope tightened before it goes out. Not for choosing a supplier, comparing priced bids (see freight-bid-comparison for freight), legal contract review, or deciding a position, which stay with the buyer.

Verified onClaude Code

Normalise a free-form purchase request into fixed intake fields while keeping the original text and turning missing fields into questions for the requester, then compare a request's recorded history to the stage requirements from requisition to payment and list the unmet conditions per stage, claimed approvals marked as unverified, and next actions with owners only where the source states them. Use when a procurement operator receives informal requests or must say where a tracked request stands and what blocks it. Not for approving spend, choosing suppliers, or navigating a specific ERP system.

Verified onClaude Code

Turn a request for real products, suppliers or manufacturers into a short verified list by writing the hard conditions first, screening every candidate against them, requiring each condition to be checked on the authority's own page with a verified date, and marking stale or unchecked claims instead of passing them. Use when someone needs three to five real candidates that meet certifications, compatibility, size, price or region limits, or a short buying brief. Not for price negotiation, supplier qualification audits, or recommending without checks.

Verified onClaude Code

Plan a reorder quantity from a sales history, lead time, MOQ and stock by treating stockout weeks as censored demand, backtesting the forecast on a holdout with a stated error, computing the order in explicit units against lead time, MOQ and pack constraints, and running demand scenarios. Use when a buyer or planner asks how much to reorder, how to turn a demand forecast into a purchase or replenishment plan, or whether a forecast can be trusted. Not for supplier selection, price negotiation, or placing the order, which stay with the buyer.

Verified onClaude Code

Help a buyer find where a procurement task lives in an SAP system or tie an SAP export to spreadsheet formulas by asking the product version and release before naming any screen or app, labelling every name as unverified until checked in the user's own system, describing change documents and approval for master-data changes, and giving a column-mapping table with a total check. Use when someone moved to a new SAP release and cannot find a purchase-order or vendor-master task, or must reconcile an export in Excel. Not for configuring SAP, authorisation design, or giving transaction codes from memory.

Verified onClaude Code

Build year-to-date and full-year spend projections and supplier spend patterns from purchase-order or transaction lines by normalising supplier names and categories first and reporting every merge, naming the projection method, labelling actual versus projected, reconciling the total to the ledger or PO total, and showing top-supplier concentration and unusual months with the rows behind them. Use when a buyer or analyst needs YTD and forecast spend sheets or supplier spend patterns from transaction data. Not for budget approval, supplier performance scoring, or tax and accounting treatment.

Verified onClaude Code

Compare free-text supplier self-assessment answers per question, list unusual or non-answering claims with the original quote, separate factual claims that need evidence from differences in wording, and draft follow-up evidence requests, without scoring or ranking any supplier. Use when a buyer has supplier questionnaire answers and wants audit follow-ups. Not for scoring, ranking or shortlisting suppliers, verifying that a supplier exists (see real-vendor-shortlist-verification), or judging a supplier's honesty.

Verified onClaude Code

Map the numbered requirements of a regulation, guidance, accreditation standard or review question list to a submission or quality document, with a status and quoted location per requirement, and list internal inconsistencies of numbers, names, versions and dates inside the document. Also covers a draft procedure or policy for using AI on regulated records, checked element by element (scope, intended use, human review and sign-off, record and audit-trail expectations, change control, training, responsibility) against the excerpts supplied. Use when someone asks whether a policy, procedure or submission meets a standard, to build a requirement-to-section table, to check a regulated document for conflicts before review, or to review an AI-use procedure. Not for legal or regulatory advice, declaring a document compliant, or approving a submission.

Verified onClaude CodeCodex

Sort a long list of regulatory review or deficiency questions by whether each applies to the product as described - applies, may not apply, or needs a decision - with the product fact and the question text quoted for every row and the assumption stated behind each "may not apply". Use when an agency or reviewer question list arrives and a regulatory lead must see which questions the supplied product description answers before assigning work. Not for deciding that a question can be ignored, for answering on the agency's behalf, or for judging the product; the regulatory affairs lead decides.

Verified onCodex

Trace a summary section of a regulatory submission (overview, clinical or quality summary) back to the source reports - split the summary into numbered statements and numbers, tie each to a document, section and quoted line, mark statements with no source found, report number mismatches in value, unit or rounding, and list source findings the summary leaves out as questions to the author. Use when a summary draft, possibly AI-assisted, must be checked against the reports it summarises before review. Not for writing the summary from nothing, judging whether it is adequate, or approving a submission; the responsible regulatory professional decides.

Verified onCodexGrok CLIAntigravity

Draft the notice to a user whose content or account was actioned from the moderation policy text and the action record only - facts relied on, rule and policy version, whether automated detection was used, the sanction checked against the enforcement ladder, and how and by when to appeal - leaving every field the record lacks blank. Use when a removal, restriction or ban needs a user-facing notice or a notice must be checked against the policy. Not for deciding the case (false-positive-appeal handles appeals), writing the policy (moderation-policy), recounting action logs, or legal advice on platform duties.

Verified onCodex

Write a moderation policy a community or platform can enforce consistently - plain rules with examples of what is and is not allowed, an enforcement ladder tied to severity and repetition, who decides and how fast, an appeal route, logging of every action, and a periodic transparency report with counts. Use when setting up or revising community guidelines, content rules for a platform, or rules for automated abuse enforcement. Not for deciding a single case.

Verified onClaude Code

Turn a moderation action log into the counts of a transparency report - recount from the rows, never from a summary column, by category, automated or human action and appeal outcome, with the denominator stated for every rate, duplicates and rows with a missing category listed, and the result reconciled against any figures already drafted. Use when an action log must become published counts or a drafted report must be checked against its log. Not for choosing what to publish, for writing the moderation policy (moderation-policy), or for general survey recounting.

Verified onCodexAntigravity

Turn an abuse signal into a detection rule that can be trusted - a precise definition, thresholds chosen on labelled data with precision and recall reported, an explanation a reviewer can check, an appeal path, and monitoring for drift - before it affects any account. Use when writing or changing a fraud, spam or fake-review detection rule, a risk score feature, or a moderation threshold. Not for brainstorming signals or handling a single report.

Verified onClaude Code

Generate and rank candidate signals for a specific abuse pattern - fake reviews, sock-puppet accounts, scraped or spun content, bought engagement - by how an abuser would behave, what data shows it, how cheaply the abuser could evade it, and what legitimate users it would also catch. Use when starting to detect a new kind of abuse or when existing rules miss a pattern. Not for writing the final rule or deciding on a ban.

Verified onClaude Code

Decide an appeal against an automated abuse action - re-read the evidence the rule recorded, check it against the rule's definition and known false-positive patterns, gather what the appellant adds, decide uphold or reverse with the reason written, restore what the action removed, and feed confirmed false positives back into the rule. Use when a user, seller or creator disputes a ban, penalty, removal or abuse label. Not for writing the rules themselves.

Verified onClaude Code

Break an unsolicited message offering a job, investment, prize or refund into its claims (who the sender says they are, any upfront fee, promised return, urgency, the action asked) and give each claim an independent check through an official channel the person finds themselves, never the link or number in the message, marking known risk signals without declaring the sender safe or a fraud. Use when someone pastes a message from an unknown sender and asks whether it is a scam. Not for moderating reports on a platform, naming a real company as fraudulent, or recovering money already lost.

Verified onClaude Code

Answer a building, electrical or plumbing code question only from code text the user pastes, after asking for the jurisdiction, adopted code, edition and local amendments, quoting section numbers only as they appear in the pasted text, separating what the text says from interpretation, and listing what the authority having jurisdiction or an engineer must confirm. Use when someone asks whether work, a dimension or an installation meets code. Not for engineering judgement, permit decisions, inspection sign-off, or contract and warranty questions.

Verified onCodexGemini CLIGrok CLIAntigravity

Split an inspection report or a homeowner warranty narrative into one defect item per finding, each with location and symptom, route each item to a trade, flag items whose trade or location is unclear, and reconcile the item count against the source findings. Use when someone pastes inspection findings, a punch list or a free-text warranty claim and wants a list sorted by trade. Not for diagnosing causes, judging severity or safety, estimating cost or deciding warranty coverage.

Verified onClaude Code

Set up the cost and scope structure of a construction package from the project scope and trade list - numbered cost codes with no overlap or gap, takeoff columns and formulas validated on a known example quantity, and a subcontract scope for one trade with its interfaces to neighbouring trades and an explicit exclusions list. Use when an estimator, cost controller or procurement lead needs cost codes, takeoff quantity formulas, or a subcontract scope and exclusions drafted. Not for pricing, rates or budgets, bid evaluation, legal contract wording, structural design, or choosing a standard code system the user did not supply.

Verified onClaude Code

Build a register of the notice, claim and time-bar duties in a construction contract and its amendments against project correspondence - trigger event, clause quoted, notice period, form and addressee, the date the trigger was known as stated in the text, the computed deadline with the day-count rule shown, status, and which amendment or change order altered the clause. Use when a contracts manager must track notice and claim deadlines from the contract and emails or letters. Not for deciding whether a notice was valid or timely in law, for event or exhibitor deadlines (deadline-register-from-documents), or for comparing contract revisions (document-version-change-impact).

Verified onCodexGemini CLIGrok CLIAntigravity

Turn a field note into requests for information (RFIs) with one question per RFI, drawing, sheet and spec references only as given, the observed conflict kept apart from a proposed answer, schedule and cost impact asked for rather than assumed, and each RFI added to the project's RFI log with number, date and status. Use when a superintendent, foreman or contractor pastes a field question or conflict and wants an RFI drafted or logged. Not for change-order pricing, contract interpretation, or deciding who is at fault.

Verified onCodex

Check a quantity takeoff or estimate sheet by writing each line's unit chain and recomputing it in code, confirming the waste factor is stated and applied once, making every total equal the sum of its lines, flagging lines far out of scale with sibling lines, and reporting each changed line with old and new value. Use when someone pastes takeoff lines, quantities or an estimate sheet and asks if the numbers are right. Not for pricing, bids or scope and cost structure.

Verified onCodex

Verify a teacher's draft test before it goes out - solve every multiple-choice item independently and compare with the answer key, test each option so exactly one is correct and every distractor is wrong for a named reason, and check that a scoring rubric has distinct observable level descriptors and weights that sum to the stated total; report every mismatch before anything is delivered and leave the teacher to decide changes. Use when a teacher asks to write, check or finalise assessment items, an answer key or a scoring rubric. Not for grading student work, assigning final grades, judging students or writing a whole course.

Verified onClaude Code

Compare an approved storyboard or source course against what an authoring tool actually built, imported or cloned - diff text and assessment items block by block and report changed wording, dropped items, reordered answers, a moved answer key and character-limit truncations with counts, and for a cloned course compare dates, access rules, grading settings and links to the source; list every difference for the owner and correct nothing silently. Use when a built or imported module may have drifted from the approved content. Not for writing the storyboard, fixing the module, judging the designer or checking learning quality.

Verified onClaude Code

Turn subject-matter-expert notes, interviews or source documents into a facilitator guide, storyboard or face-to-face activity without inventing content - tag every statement with its source passage, list unsourced points as gaps with a question for the expert, add a time estimate per segment summed against the session length, list the prerequisite knowledge the source does not give, and keep the template's sections in order. Use when a course draft, facilitator guide, storyboard or activity must be built from SME material under a template and session length. Not for writing assessment items, grading rubrics or choosing reading levels.

Verified onClaude Code

Draft measurable IEP goals from a supplied baseline, with condition, behavior, criterion, measure and timeframe, plus home-support suggestions and practice items traceable to each goal, all marked as drafts for the team to review. Use when a special education teacher or case manager is writing or adjusting a goal, a parent home-support suggestion or matched practice items for one student. Not for diagnosis, placement or eligibility decisions, legal compliance review, or any request that includes a real student name or identifier.

Verified onClaude Code

Run language practice that stays in one variety - translation drills one item at a time with correction after the learner answers and an error log, near-synonym contrasts with example sentences, and dialect-controlled examples. Use when a learner wants translation drills, oral or written practice, the difference between two similar words, or examples in a specific dialect or register. Not for certified translation or exam grading.

Verified onClaude Code

Fit a lecture script and slide deck to the allotted time - convert script word counts to spoken minutes at a stated words-per-minute rate, add activity, poll and question time, show minutes per slide per section against the total, flag the overrun and the slides that run long, and name concrete cuts or moves to speaker notes that close the gap. Use when a lecture's script and slides do not fit the time and the user has been trimming repeatedly. Not for structuring the story of a deck (see deck-storyline), for reviewing slide content or for judging a speaker.

Verified onClaude Code

Draft a single lesson plan that provably meets its fixed constraints - activity minutes summed against the period with the sum shown, every activity mapped to a stated objective, a named accommodation placed in a specific activity for each stated learner need, any caps or fixed slots (teacher talk limit, exit ticket last) checked, and a materials list. Use when a teacher gives a period length, objectives and learner supports and wants a plan. Not for multi-session course design (see course-draft-from-sme-material), for judging a student, for diagnosing needs or for replacing the teacher's judgement.

Verified onClaude Code

Turn pasted study notes into a small deck of atomic recall flashcards plus a spaced review schedule that ends before the exam - one testable fact per card, each card traceable to a line of the notes, no card that only restates its own answer, trivia skipped and listed, the deck capped and prioritised, and every card checked against the notes. Use when someone wants flashcards or a review plan from their own notes. Not for inventing facts the notes lack, replacing the course material, judging a learner or generating large decks.

Verified onClaude Code

Write or adapt a reading passage, warm-up or practice text to a stated grade level by computing a readability formula and revising until the measured level is within one grade of the target, while keeping facts accurate and building on the learner's interest. Use when a teacher or tutor asks for a passage at a grade level, to simplify a text for a grade, or for fluency material on a student's interest. Not for diagnosing a student or writing assessments for placement.

Verified onClaude Code

Check a draft or analysis against the numbered requirements of an assignment brief, rubric or research question - extract the requirements before reading the draft, map each to quoted evidence with its location or mark it uncovered, report partial coverage as its own category, and leave the draft unchanged. Use when someone wants to know whether their paper, report or analysis answers everything the brief asks, or which requirement they missed. Not for rewriting or improving the draft, grading it, scoring quality, detecting plagiarism or judging the author.

Verified onClaude Code

Produce word searches, grid puzzles or short music notation by writing code or notation text instead of asking an image model, verify every target word at its listed coordinates or every bar against the time signature with the bundled checker, give the answer key as coordinates or a playable text form, report any failure, and render to a print format last. Use when a teacher or parent wants a printable word search, similar rule-bound puzzle or simple melody and image generation would put words or notes in the wrong places. Not for decorative images, for checking an AI image (see generated-asset-delivery-spec-check), for composing music of artistic quality or for GPU generation.

Verified onClaude Code

Trace SCORM completion/success, API returns, persisted state and relaunch observations for a specific SCO and attempt. Use when final slides or scores disagree with LMS status or resume state. Not for grading learners, forging completion or inferring edition-specific behavior without a supplied contract.

Verified onCodex

Act as an independent second reader on student answers that a teacher or assistant already scored - grade each answer blind against the rubric before looking at the given score, quote the answer text that supports each rubric point or shows its absence, rank disagreements by score gap and list the top items for the teacher to recheck, flag rubric ambiguity, and never give a final grade. Use when a teacher wants to know which scored answers to recheck, or wants a second look for points missed on first reading. Not for assigning final grades, judging a student, detecting cheating or writing the rubric.

Verified onClaude Code

Draft a parent message or referral note about a difficult student situation from the facts the teacher supplies - observed behaviours with dates, supports already tried and their result, a clear ask and next step - in neutral wording with no labels, diagnoses, causes or character judgments, and with a short list of what was left out or not supplied. Use when a teacher or school staff member asks for help writing to a guardian, a counsellor or a referral team about a student concern. Not for diagnosing or labelling a student, predicting outcomes, deciding a referral, or adding facts the teacher did not give.

Verified onClaude Code

Spread a textbook or course over the available school weeks - count weeks and study hours after stated breaks, total the effort per unit against that capacity and state the shortfall or slack, show the effect of dropping review chapters or appendices as a choice for the user, place subjects shared by several children in the same slot with the parent's supervision free of clashes, and show what slips if a week is missed. Use when a parent or self-learner has a book with no pacing and a number of weeks. Not for a single day plan (see constrained-day-schedule), for judging a learner or for deciding what a child must study.

Verified onClaude Code

Audit a pasted worked solution one step at a time - rewrite it as numbered equivalences with the rule used and its conditions, find and quote the first incorrect step, show the corrected step, redo only what follows, and verify the final result by an independent method such as differentiating, substituting back or a numeric spot check; for a textbook-style explanation, tabulate the claims against the pasted text and mark causal claims it does not support. Use when a student pastes their own solution or an AI explanation and asks where it went wrong, whether it can be trusted, or wants every step shown. Not for solving a graded problem the student has not attempted, for writing answers to hand in or for judging a student.

Verified onClaude Code

Compile part numbers, instruments or venues into a specification table where every value is quoted with the maker or source document and its date, unverified cells are marked instead of filled from memory, units and the conditions a figure holds under are kept, and price or availability is left as to be checked. Use when someone asks for package types, ratings, bandwidth, capacity or similar specs for a list of parts or models. Not for choosing a part, for a bill of materials with prices, or for a standards check (see standards-clause-locator).

Verified onClaude Code

Convert a pasted register or tag table into the import file or device map a target format asks for, row by row - state the address base and conversion, register width and data type per row, check unique names and overlapping or non-contiguous ranges, keep an input and output row count, list every value the table does not give and every assumption, flag rows that do not fit instead of coercing them, and say the file must be tested on a bench device before it reaches equipment. Use when an engineer pastes a Modbus register table, PLC tag list or device map and wants a configuration or import file in a given format. Not for extracting specs from a datasheet or for commissioning real equipment.

Verified onClaude Code

Turn a manufacturer manual excerpt and nameplate into an asset register entry and a preventive maintenance schedule where every interval and trigger is copied from the manual with its page, hours-based and calendar-based triggers stay separate, missing fields are listed for site capture and no interval is inferred. Use when a facilities or maintenance person pastes manual pages and nameplate data for new equipment. Not for choosing intervals the manual does not give, for finding which standard applies (see standards-clause-locator), or for approving the plan, which a qualified technician does.

Verified onClaude Code

Compare SVG or DXF physical geometry and approved cut, engrave and non-output roles with recipient import observations. Use when CAD layers collapse, vector sizes change or a browser preview is offered as fabrication proof. Not for machine commands, power settings or safety certification.

Verified onCodex

Take a vehicle, HVAC or similar equipment symptom and produce the five to eight questions that best separate the possible causes, in order of diagnostic value, plus a running evidence log of symptom, test, reading with units, result and what it rules out, keeping observed facts apart from guesses and suggesting only safe next checks. Use when someone describes a machine that will not start, cool or run right and wants to know what to check and how to record it. Not for naming a root cause, for software defects (see defect-triage-evidence), or for repairs, which a qualified technician decides.

Verified onClaude Code

Draft a job hazard analysis or work-order risk packet from the task steps and site rules the user supplies - a step-by-step table of hazards, controls, residual risk and PPE, a standards or site-policy column that cites only what was supplied and flags the rest as to be confirmed, permit and lockout triggers, special-environment triggers such as infection control in an occupied healthcare area or confined space, and every control marked for a competent person to confirm. Use when someone asks for a job hazard analysis, JHA, work-order risk assessment or pre-task hazard review. Not for declaring a task safe, approving work, setting exposure limits or ratings, or replacing the safety officer's site review; for finding a clause in a standard use standards-clause-locator.

Verified onClaude Code

Turn field steps, inspection notes, a technician report or a migration plan into a work instruction, process flow or cutover plan that keeps every step and branch as stated, gives each step an owner and prerequisites, marks gaps as questions and lists what the performer and safety owner must verify. Use when someone pastes rough operating steps and asks for a work instruction, procedure, flow or runbook draft. Not for inventing safety steps or values, for a live incident runbook (see operator-runbook) or for approving a procedure, which stays with the responsible engineer.

Verified onClaude Code

Extract hazard classification, composition and precaution fields from safety data sheets by section number with page and quote, record supplier and revision date per sheet, refuse to merge sheets of different versions, and list every change between a supplier's old and new sheet as old versus new with page references; blank or unreadable fields are reported as not stated. Use when someone asks for a hazard and composition table from sheets in mixed layouts or what changed in a supplier's revision. Not for classifying a substance, setting exposure limits, approving a product, or any safety-compliance decision.

Verified onClaude Code

Find which clause of a supplied code, specification, standard or equipment manual governs a stated requirement or symptom, quote the clause number and wording verbatim with its edition, compare report values against clause values by number, unit and condition, and say plainly when no clause was found. Use when an engineer pastes a report, spec excerpt or manual section and asks which clause applies or whether two documents agree. Not for supplying clause text from memory, for deciding compliance, or for building a maintenance schedule.

Verified onClaude Code

Compare CAD assembly occurrences, placements and approved physical dimensions with receiving STEP observations. Use when exported geometry looks correct but scale, linked instances or required metadata differ. Not for constructing CAD, validating kinematics or certifying manufacture.

Verified onCodex

Build or re-plan a one-day schedule - a run-of-show, a travel day, a family outing, a tour or visit round, a move or a multi-dish cook - under fixed bookings, durations, travel time, opening hours and protected blocks such as a nap, with a worked timeline, a check of every item against its hours, buffers and backups for what could change. Use when someone asks to plan or re-plan a day or event around fixed times, opening hours, travel between stops, a nap or rest window, a closed venue or a change that hits the plan. Not for multi-week calendars, booking or buying anything, or live availability the user did not supply.

Verified onClaude Code

Build a deadline register from pasted event or exhibitor documents - each item with its date, time zone, owner and the quoted source line - and diff a new version against the previous one. Use when someone pastes a kit, guide or schedule (often revised) and asks for a list of deadlines, who owns each, or what changed between versions. Not for calendar invites, reminders, legal or contract deadlines, or documents the user has not pasted.

Verified onClaude Code

Compare recurring iCalendar source identities, value types and time zones with bounded expected and observed imported occurrences. Use when an ICS import loses exclusions, shifts recurring times or mishandles a moved occurrence. Not for planning schedules, booking events or certifying an importer from parse success alone.

Plan or audit a transit commute or a hiking route under stated closures by checking every segment, station and transfer against the official sources the user supplied, excluding closed lines, and marking everything unchecked as unverified. Use when someone asks for a route, a commute around a closed line, or a hike and the answer could name wrong stations, closed lines or misplaced mountains. Not for live delays, timetables, fares or entry rules the user did not paste, or for presenting an unchecked route as an instruction.

Verified onClaude Code

Seat guests at tables from the keep-apart, sit-together and capacity rules the user supplies, then verify every rule against the final layout and report the conflicts instead of bending a rule. Use when someone asks to make a seating chart, table plan or room split for an event, dinner or class under rules such as A apart from B, these people together, N seats per table. Not for guessing who gets along, ranking or judging guests, venue floor plans, or invitations.

Verified onClaude Code

Build a hotel or trip shortlist for a traveler's stated conditions from an approved list the user supplies, with each condition turned into a check, each option scored per check from the supplied facts, and availability and current price marked as still to confirm. Use when someone pastes a list of candidate hotels or trip options and a traveler's needs (family, distance to transit, kitchen, budget) and asks for a shortlist or comparison. Not for finding properties outside the supplied list, booking, live prices or availability, or ranking by review score alone.

Verified onClaude Code

Compute a numerology reduction, four-pillar (saju) or natal chart input, tarot spread or dream-symbol reading only with the rule set and tables the user names or supplies, showing inputs and every step, keeping tradition-specific rules such as master numbers exactly as stated, flagging steps another system would change, and framing the result as a tradition-based reading for reflection with no prediction or advice on health, money or relationships. Use when a user asks for a life path number, a chart or pillars from a birth date, a spread reading or a dream reading and the answer depends on calculated values or their own symbols. Not for forecasts, medical, financial or relationship decisions, or inventing a table the user did not supply.

Verified onClaude Code

Parse a user's log of tarot or dream readings into date, question, cards or symbols and reading, count recurring cards, symbols and repeated questions by exact tally, report only patterns the counts support, flag readings that merely echo the answer the user said they hoped for, and flag when the user plans to decide something on the strength of a reading, answering with reflection prompts rather than outcomes. Use when someone pastes a journal of readings or dreams and asks what patterns recur or what the cards keep saying. Not for predictions, for telling the user what to decide, or for advice on health, money or relationships.

Verified onCodex

Read a tarot spread so that each card is tied to its named position and to the user's question, with the draw recorded first as a table of position, card and orientation, and with the draw coming only from the user's own physical draw or from a random method shown step by step, never from cards picked to fit the topic. Links adjacent cards, keeps the user's own card associations ahead of generic text, and frames the reading as reflection. Use when someone asks for a spread such as past, present, future or a celtic cross to be read, or asks to be dealt cards. Not for numerology, birth chart or pillar arithmetic (divination-calculation-fidelity), predictions, or advice on health, money, legal or relationship decisions.

Verified onCodexGemini CLIGrok CLIAntigravity

Isolate a home audio-video fault such as a soundbar that loses eARC or ARC sound when the TV sleeps, or a first room measurement that looks wrong - ask for models, firmware, ports and the real wiring, draw the chain, change one thing per step with the expected result for each, separate audio transport from control (CEC) problems, and for room measurement check microphone calibration, position and repeat runs before reading the graph. Use when a user keeps resetting a TV, player or soundbar, or doubts a speaker measurement. Not for generating settings from a photo, applying one user's workaround to every product, or electrical repair.

Verified onClaude Code

Organise a person's own journal or mood entries into notes for their next session with a counsellor or therapist - dated entries with their own words quoted, counts of the indicators they chose (and only those), and questions for the professional each pointing to the entry that raised it - without interpreting causes, diagnosing, or calling a trend better or worse. Use when a user pastes their own log and asks to prepare for a session, or asks whether their logged scores are improving. Not for therapy, diagnosis, crisis support or analysing why someone feels as they do.

Verified onClaude Code

Keep an assistant inside the kind of help the user chose when they talk through a hard time - record the scope (listen only, information only, reflective questions, or a step list they already learned), restate it in one line, ask permission before any change of mode, recite only the steps the user supplied, and never diagnose, interpret or reframe. Use when a user wants to be heard, says not to analyze or reframe them, complains the assistant switched into advice or correction, supplies a coping list to be read back, or sets a rule for flagging repeated worry questions. Not for therapy, crisis counselling, clinical advice or analysing a person's feelings.

Verified onClaude Code

Check calorie and macro targets and meal plans by arithmetic - state the named formula (for example Mifflin-St Jeor) with every input and the activity multiplier, compute the energy target in code, set protein from the user's own rule, rebuild each meal's and each day's calories from grams at 4, 4 and 9 kcal per gram, compare with the label or claimed figure and with the target, and flag any day more than 5 percent off. Use when someone asks for calorie or macro targets, or pastes a meal plan or food log with totals to be checked. Not for medical diets, eating disorders, pregnancy, children or clinical nutrition, which are referred to a clinician or dietitian.

Verified onCodex

Check a running or race plan as a week table of total distance, long run and quality sessions, flag weekly distance jumps above the user's stated cap or the general 10 percent rule quoted as a rule of thumb, the long run's share of the week, missing cutback weeks, the taper length and a peak long run that is short for the race distance, and set paces only from a dated recent result by a named method. Use when someone pastes or asks for a running, half marathon or marathon plan to be checked, or wants paces for it. Not for injury or pain decisions, which go to a qualified professional, or for guaranteeing a finish time.

Verified onCodexGrok CLI

Turn a pasted activity or recovery export (CSV or a typed summary) into a weekly table where measured values (distance, duration, heart rate) are summed or averaged by stated rules and kept apart from the vendor's composite scores (readiness, training status, race predictor), which are shown in their own column labelled as vendor estimates, reporting trends only when there are at least three weeks of data and drawing no injury or medical inference. Use when someone pastes watch, ring or app data and asks for a weekly summary, a trend, or what a score means for their training. Not for injury risk, medical readings of heart rate or heart rate variability, or writing a training plan.

Verified onCodex

Structure a pasted workout log into a sets, reps and load table with per-lift changes, propose equipment substitutions by movement pattern for the user's available equipment, and mark every load change and substitute as a suggestion for the user's trainer or coach to approve. Use when someone pastes training logs and wants a progression summary or a plan adapted to home or gym equipment. Not for injury, pain or medical advice, diet, writing a new programme from scratch or replacing a trainer.

Verified onClaude Code

Check a hazard analysis or HACCP plan draft, AI-written or not, against the process flow - a step table of hazards and controls, completeness of every named control point (critical limit, monitoring, frequency, corrective action, verification, record), a pasted source for every number, and flow steps the draft never mentions listed as questions. Use when a food business or QA lead has a HACCP plan draft and a process flow to review before the HACCP team meets. Not for deciding which step is a critical control point, setting limits, or claiming a plan meets any regulation or will pass an inspection.

Verified onCodex

Build a staff lookup table from a menu's approved ingredient lists and supplier label statements, flag cross-contact points for staff, and write a training quiz only from the confirmed menu, listing allergens only as the supplied labels state them and leaving the decision to a qualified person. Use when a restaurant manager or trainer wants an allergen and ingredient reference table or a new-staff menu quiz. Not for declaring any dish safe or free of an allergen for a guest, for creating recipes, or for legal labelling advice.

Verified onClaude Code

Compare a label draft for prepacked food sold on the premises - name, ingredient list, allergen emphasis, any may-contain line - with the approved recipe and each supplier label statement, including allergens inside compound ingredients, and mark every allergen missing, extra, unclear or unconfirmed. Use when a food business has a label draft and the recipe and supplier statements to check it against. Not for declaring a product allergen-free, for law-specific labelling advice, or for the staff lookup table and quiz (menu-allergen-index-quiz, a separate skill).

Verified onCodexGemini CLIGrok CLIAntigravity

Turn a pasted recipe into a card that keeps every ingredient, quantity, temperature and time exactly as written, then log each adaptation (an ingredient swap, a different appliance) as a separate marked change with the original value, the new value, the function the item served and a doneness check, with source numbers kept apart from estimates. Use when someone pastes a recipe and asks for a recipe card or a tidy version, asks what to use for a missing ingredient, or wants a dish moved to a slow cooker, air fryer or another appliance. Not for scaling a recipe for yield and cost, menu planning, nutrition counts, or inventing a recipe from a dish name.

Verified onClaude CodeCodexGrok CLI

Scale a recipe from its base yield to a target number of portions and compute cost per portion, applying trim loss in the right direction, converting units explicitly and costing from the purchase unit price, with the chef or manager confirming loss rates and prices. Use when a kitchen or caterer needs a recipe for 50 or 120 portions, an ingredient purchase quantity, or a menu item cost. Not for copying or adapting a recipe (use recipe-source-fidelity-and-adaptation), menu planning, or nutrition values.

Verified onCodexGrok CLI

Plan a multi-day menu starting from the stock on hand and its use-by dates, compute what must still be bought as quantities (needed minus on hand), and check storage and feasibility, with the cook confirming dates, amounts and fit. Use when someone gives a stock list, a number of days and eaters and wants a menu that uses what they have and a minimal shopping list. Not for recipe adaptation, costing a recipe, nutrition planning, or judging whether old food is safe to eat.

Check family-tree exports for impossible relationships and dates (a parent too young or born after the child, a death before a birth, an event after death), compare two exports by identifier and list differing names, places, relations and dates without merging or correcting, and build a per-ancestor sheet that keeps direct evidence, interpretation and open questions in separate labelled sections. Use when a user has tree exports or ancestor notes and wants errors found before an import, or a fact sheet for one ancestor. Not for tracing living people, confirming a relationship, or editing the user's data.

Verified onClaude Code

Plan where to search next for a named person, place and period - separate the historical jurisdictions (township, civil county, parish, diocese and their predecessors) from the nearest named place, match record series to coverage, date range, surviving volumes and custodian from the finding aids supplied, tell indexes from images and online from onsite, and return an ordered search plan with a log that keeps "not searched" apart from a documented negative. Use when a family historian asks which records, archive or collection to try next. Not for confirming ancestry, transcribing handwriting, checking tree exports, locating living people or deciding citizenship or inheritance rights.

Verified onCodex

Index every mention of a surname across local-history volumes or OCR text with volume, page, line and surrounding context, flag same-name different-person risk and spelling variants, list unreadable or unscanned pages as gaps instead of as no mention, and offer candidate readings of damaged handwriting with the comparison letters used, every reading marked uncertain until checked against the image. Use when a researcher searches long family or local-history volumes for a name or asks what a damaged line of handwriting says. Not for confirming a relationship, proving identity, or legal document OCR risk review.

Verified onClaude Code

Check or write a crochet or knitting pattern by recomputing every round or row stitch count from the stated increases and decreases with a running total, listing every part with its join point and attach step against the requested shape, and keep photo diagnoses and garment alteration advice to ranked causes and reversible steps. Use when a crafter asks for a crochet or knit pattern for a shape, wants an AI-written pattern checked before stitching, asks what is wrong with a stitch pattern in a photo, or asks how to alter a garment. Not for choosing yarn brands, selling patterns or producing a pattern image.

Verified onClaude Code

Play or referee a board game, chess or a family tabletop role-playing session with the model as opponent, rules lookup or game master by keeping an explicit state block (position or resources, whose turn, history) that is rebuilt from the move list and re-printed after every move, rejecting any move that is not legal from that state with the rule or line cited, answering rules questions only from the pasted rulebook text and edition, and ending a role-playing session with a resume note. Use when a user wants to play chess or a board game against the model, asks what a rule says, or runs a family tabletop session and the model keeps losing the board, inventing moves or mixing editions. Not for generating puzzles, teaching strategy or choosing which game to buy.

Verified onClaude Code

Plan a garden bed or small farm from measured site facts and check the plan against them - one site brief (dimensions, sun hours, soil pH and texture, last and first frost, water test with units and method), bed layout checked for row count, spacing and area used and left over, a sowing and harvest calendar built from the stated frost dates and variety days with the source of each interval, and plant suggestions checked against light, soil and native range with guesses marked. Use when a grower gives bed size, sun, soil, frost dates or a water test and asks for a layout, calendar, plant list or what to confirm. Not for pest or disease diagnosis, or for guaranteeing yields.

Plan a garden or farm chemical application only from the product label text the user pastes - quote the rate line and crop listing, convert the rate to the user's area and sprayer volume in code with units shown, give the pre-harvest interval, the re-entry interval, the maximum applications and the minimum days between them, check the planned spray and harvest dates against them, allow a tank mix only where the label says so, and mark anything the label does not cover as not on the label, to be asked of the manufacturer or a local extension office. Use when a grower pastes a pesticide, herbicide or fungicide label and asks how much to mix, when to spray or harvest, or whether to combine products. Not for choosing a product, for off-label or mixed uses the label does not state, or pest and plant identification (plant-photo-id-confidence-gate).

Verified onCodexGemini CLIGrok CLI

Handle a request to identify a plant, pest or plant disease from a photo or a description of one by listing only the visible features, ranking two or more candidates with the feature that supports and the feature that argues against each, always naming the dangerous lookalikes for the family (for example poison hemlock against wild carrot, parsley or other umbellifers), asking for the extra photos that would settle it, and sending any edibility or toxicity question to a poison control centre or a local extension service without ever calling the plant safe or edible. Use when a grower or walker asks what a plant, bug or leaf problem is from a picture or description. Not for foraging approval, pet, livestock or medical advice, or garden planning (garden-plan-constraint-check).

Verified onCodex

Answer a question about a cosmetic or household product's ingredients or layering compatibility from the label text the user pastes, not from the product name, separating what the user observed from what is inference, pointing to an ingredient reference the user can check, and proposing a small trial instead of a verdict. Use when someone asks whether two products work together, why a product pills or balls up, or what is in a product, and gives a name or photo description. Not for medical or skin-condition advice, safety or allergy guarantees, or efficacy claims.

Compare bundled personal offers, such as a car price with a trade-in discount against selling the old car separately, or a pay-per-GB mobile plan against a flat plan, by splitting each offer into its written price components with validity dates, recomputing every total with the arithmetic shown, running the plans against the person's own usage figures, and marking each rate, fee or rule that was not in the pasted quotes as unverified. Use when someone asks which of two offers costs less in total. Not for choosing for the person, for financial, tax or legal advice, for business bids, or for checking network coverage.

Verified onClaude Code

Turn pasted service emails, appointment notes or family messages into a dated ledger with one row per event, classified completed, scheduled or cancelled, each row quoting its source line, with unknown dates and reasons kept unknown, conflicting records shown side by side, and next-due dates computed only from an interval the user or manufacturer text states, with the calculation shown. Use when someone asks what was done when, what is due next, or why something was bought, from records they paste. Not for extracting commitments to act on (email-history-commitment-mining), for deciding what maintenance is needed, or for medical, legal or financial conclusions.

Verified onClaude Code

Reconcile a policyholder's own paperwork - match bills, explanations of benefits and payment proofs by date of service and claim or invoice number, show billed, allowed, insurer paid, patient owed and patient paid per claim with who received each payment, list unmatched items, and order insurer letters by date to tabulate reasons and flag changed reasons or evidence submitted but not mentioned - with a source for every row. Use when someone pastes medical bills, EOBs, payment records or months of insurer letters and asks who was paid what or how the insurer's statements changed. Not for building an appeal, coverage or legal conclusions, or telling the insurer what it owes.

Verified onClaude Code

Turn a person's own numbers - pay dates and net pay, bills with due dates, debts with balance, rate and minimum, dated one-off or yearly expenses, or a mixed receipt - into a table per pay date that places every item by its due date, never assigns a dollar twice, recomputes interest with a stated formula, sets aside future expenses per pay period marked estimated or confirmed, and splits a receipt by line item so the parts sum to the total including tax. Use when someone wants a pay-period plan, an extra-payment allocation or a budget split from figures they supply. Not for financial, debt, tax or credit advice, for choosing which debt or product to use or for verifying loan terms.

Verified onClaude CodeGemini CLI

Turn a cluttered space, a seasonal changeover or recurring home and vehicle upkeep into a short sequence of physical actions that fits the time the person has, starting with the smallest set of actions that makes the space usable, each action with the items it needs, the place they move to and a time estimate the person can correct, and dated so every deadline they stated is met. Use when someone asks what to do first, how to start, or how to spread chores over days or weeks. Not for scheduling a single day by the hour (constrained-day-schedule), for health, mental-health or safety advice, or for judging how the person lives.

Verified onClaude Code

Audit an individual's own trade journal against the trading rules they wrote - tabulate each trade against each rule as pass, fail or not stated with the journal line quoted, group failures by rule with dates and amounts and no outcome speculation, flag missing fields and hindsight wording in notes, and state assumptions such as market open time. Use when a user pastes a trade journal and their written rules and asks for deviations or similar past situations. Not for recommending trades, predicting prices, judging whether a strategy works, or promising results.

Verified onClaude Code

Draft a batch of personalised donor thank-yous or stewardship updates from approved giving records, relationship notes and approved program facts - reconcile donor and gift identifiers first, separate gifts from pledges, sponsorships and volunteering, build a short brief per recipient, draft by segment while keeping each person's own details, mark missing context as placeholders, then reconcile drafts against the recipient list and return an approval queue with exceptions for the development lead. Use when a development team writes many thank-yous or updates at once from a gift list. Not for grant applications, funder selection, wealth inference, tax acknowledgments or receipts, invented beneficiary stories, or sending anything.

Verified onCodexGrok CLIAntigravity

Check which funders or grant calls fit an organisation by comparing the organisation's own facts with the call text the user pastes - one row per funder with the eligibility and deadline sentences quoted from the call, a fit verdict of meets, does not meet or cannot tell for each criterion, the deadline compared with today's date, stale or mismatched listings flagged, and funders with no call text marked unverified. Use when someone has a list of possible funders, often from a search or an AI summary, and asks which are open to their organisation. Not for finding funders without call text, predicting funding odds, writing the application, or deciding eligibility, which the funder's programme officer does.

Verified onClaude Code

Fit a grant response to the funder's questions and limits - list each question with its word limit first, cut or restructure a draft to the limit while keeping the required points, report the final word count, keep activities apart from outcomes, exclude unconfirmed numbers, and report which questions or required reference points are not covered. Use when someone adapts an existing answer to a new word limit, outlines a response from a project memo, or asks what their draft misses against the funder's questions or toolkit. Not for deciding what the funder wants, promising funding odds, or inventing figures.

Verified onClaude Code

Abstract a commercial lease into a fixed set of key terms - parties, premises, term, options, rent schedule, escalations, CAM and expenses, deposit, renewal, termination, assignment - each with clause number and quote, recompute rent and escalations independently and flag mismatches with the lease's own figures, convert relative dates to calendar dates from a stated reference date, and mark absent terms NOT FOUND. Use when a lease administrator, broker or analyst needs an abstract or a critical-date list from lease text. Not for legal advice, negotiating terms, or filling gaps with what is standard.

Verified onClaude Code

Answer whether a parcel or proposed use meets a municipal bylaw or zoning text by quoting the governing sections, splitting the question into numeric conditions, testing each with the arithmetic shown, and reporting effective date and overlay or variance clauses. Use when a user supplies or asks about local ordinance text for a lot, building footprint or use, or screens a site's early feasibility. Not for a legal opinion, a permit decision, or answering from memory when the text was not supplied.

Verified onClaude Code

Compare home listings against a buyer's requirements by sorting them into must-have, trade-off and nice-to-have, marking each as met, not met or UNKNOWN from the listing text only, and preparing a per-property viewing sheet with questions to verify on site. Use when a buyer or agent gives requirements and several listings or candidate addresses. Not for valuation, price negotiation, or recommending a purchase.

Verified onClaude Code

Plan a furniture layout, render prompt or room description from a floor plan or measurements by first writing a fixed-structure table (walls, windows, doors with swing, units, each value marked given, derived or not given), placing furniture only from stated dimensions with clearance arithmetic shown, and listing every wall, opening or object that moved, appeared or vanished when a layout or generated image is compared back. Use when a user gives room dimensions, a plan or photos and wants furniture placed, a walkway checked, or a render checked against the real room. Not for ordinance or zoning checks, structural or safety advice, or inferring dimensions from one photo with no reference.

Verified onClaude Code

Organise a deposit or small-claims dispute into an evidence packet from the user's own records - a dated timeline from lease, move-in and move-out documents, an item-by-item condition comparison with document references, each amount claimed tied to a lease clause and receipt with ordinary wear and age flagged, deposit and notice deadlines counted from the dates given and marked met, at risk or unknown, and a list of missing evidence. Use when a tenant, landlord or property manager has the documents for a deposit or move-out dispute and wants them arranged for an attorney, tenant advisor or the court clerk. Not for legal conclusions, who wins, deadlines or limits of a jurisdiction not supplied, drafting the legal claim, or contacting anyone.

Verified onClaude Code

Audit the permission and approval setup of a coding or browser agent that runs unattended - walk every allow, deny and ask rule to see whether a run with no approver stops, skips or exits zero looking successful, treat a denied command and any variant with changed flags, paths or wrappers as the same denied request, keep credentials out of model input and logs, and verify the expected file actually changed. Use when someone runs an agent nightly or in CI, an agent retried a denied action with a bypass flag, a run finished with no edits, or an agent hit a login wall, and they ask whether the settings are safe and will complete. Not for connector or MCP server review, or for writing the agent's task prompt.

Verified onClaude Code

Produce the data-flow document a security review asks for, per skill or MCP connector - data read and written, where it is sent, what is stored and where, retention, credentials used and their scope, and network egress - with every line marked measured (observed in code or traffic) or declared (stated by the vendor or author). Use when a connector, plugin or skill was rejected or is pending for lack of a data-flow description, when a security questionnaire asks where data goes, or before approving a new MCP server. Not for a privacy policy for end users or a full threat model.

Verified onClaude Code

Inventory which MCP servers and tools each person has connected, read from each agent client's config files, compare them with the company allowlist, and report unapproved, unknown-version and over-permissioned connections per person, with a suggestion to measure real client traffic through a logging proxy. Use when the user asks which MCP servers people have connected, wants to audit agent tool connections against an approved list, or suspects unapproved connectors in use. Not for building an MCP server or for reviewing one server's data flow.

Verified onClaude Code

List everything a departing member holds across agent tools - skills and plugins they own or published, MCP connectors and the tokens behind them, API keys, service accounts, shared prompts and assistants, and approval rights - then produce a revocation checklist with an owner, a handover target and confirmation evidence for each item. Use when someone leaves the team or changes role and their AI tool access, keys or approval rights must be reviewed, transferred or revoked. Not for general HR offboarding or for auditing everyone's MCP connections.

Verified onClaude Code

Diff two versions of an agent skill or plugin and flag every change that needs re-approval - tools and permissions allowed, external connections and hosts, data paths read or written, scripts and hooks added or changed, and dependency changes - then produce an approval checklist bound to the exact version hashes reviewed. Use when someone asks whether a new version of a skill, plugin or MCP server can be approved or rolled out (especially on the strength of a changelog), when a security or engineering reviewer must approve such an update, when the code that was reviewed may differ from the code that runs, or before rolling a new skill version out to a team. Not for a first full security audit of an application or for writing the skill.

Verified onClaude Code

Respond correctly when a guard hook refuses a shell command (sed -i, git push --no-verify, docker compose down) - read the refusal, take the stated alternative, never re-spell the command to get around it, and report a real false positive with the exact command. Use when a Bash command was blocked or refused by a hook, when asked to bypass a guard, to push with --no-verify, or to edit in place with sed, perl or awk, or when a guard seems not to be working. Not for writing new guard rules or for approval prompts that are not a hook refusal.

Decide whether a task goes to zero, one or several subagents and how many - classify it as wide read-only, write-coupled or sequential, keep small or sequential work in one agent and say why, size a fan-out by breadth with a stated token multiple, pick the mechanism (turn-by-turn helpers, a script, a peer team) by who holds the plan, and set caps and a pilot slice before launch. Use when a user asks to use agents, spin up a team, parallelize or fan out work, or when you are about to spawn subagents. Not for writing the subagent prompt itself or for scripting a large run.

Verified onClaude CodeCodexGemini CLIGrok CLI

Keep a with-versus-without measurement honest when the change is a skill, prompt or agent and the graders are models - hold subject model, judge, grader text, prompt and environment constant across both arms run in the same batch, invalidate old baselines after any grader or prompt edit, keep references independent of the test case, use a judge strong enough to reject a known-bad answer, rerun an unexplained negative verdict once, cap fix rounds and reserve the final run budget. Use when reading, reviewing or planning an eval log, a before-after score or an uplift claim. Not for production monitoring dashboards or for verifying a single subagent report.

Verified onClaude Code

Design a script that drives dozens or more agents - phases of discover, per-item pipeline, independent verification and one reduce step, barriers only where a stage needs every result, a deterministic script with no clock or randomness and per-item prompts independent of siblings so a resume reruns only what changed, schema hand-backs with explicit handling of failed agents, caps on heavy local tools and agents, a pilot slice and loop stop rules. Use when asked to outline or write a workflow script, batch job or pipeline over many files or items with agents. Not for a few turn-by-turn helpers.

Verified onClaude Code

Plan several agents that write to the same repository or store - list every path and shared record, give each to exactly one agent, isolate writers in separate copies while naming what isolation does not cover (ports, databases, caches, the shared index), have workers write per-agent pending files, and let the coordinator merge them serially with an idempotent compare-and-set that skips existing entries and applies an update only when the current value matches the recorded one. Use when more than one agent or session will edit files, a registry, an index or version numbers at the same time. Not for read-only fan-outs.

Verified onClaude Code

Write the prompt for a subagent so it can succeed alone - one-sentence objective and done-state, the context it cannot see, the exact write set it owns and what it must not touch, a value contract for every output field, a stop rule with a budget, forbidden shortcuts with an explicit blocked exit, and a hand-back schema that separates done, evidence, not verified and open. Use when asked to write, draft, paste-ready or review a prompt, task or instructions for a subagent, worker agent or delegated task, or to say what it should send back. Not for deciding whether to delegate or for splitting a uniform job into chunks with one shared brief file.

Verified onClaude CodeCodexGemini CLIGrok CLIAntigravity

Use when the user pastes a subagent's or worker's report claiming results (all tests pass, 5 of 5 cases at 1.00, fixed, only touched one file) together with its output, and asks you to draft the announcement, say whether it is safe to accept, merge or ship, or verify it first. Do not take the report at its word - list the claims that matter, re-check each against the pasted artifacts (result tables, git status, timestamps, models used), look for shortcuts such as edited tests or graders, and answer with verified, refuted and unverified claims kept apart. Not for designing an evaluation or comparison.

Verified onClaude CodeGemini CLI

Decide what belongs in the system prompt of an agent that calls tools, and what to leave out, with the symptom each omission produces. Use when writing or reviewing a tool-using agent's instructions, or when an agent misuses its own tools.

Verified onClaude Code

Cut a package's thin rules file (AGENTS.md or CLAUDE.md) under its byte budget by keeping only constraints there and moving everything else verbatim to the long-form document, settling any contradiction between the two by measuring the code, and keeping links from the thin file to the long form. Use when an AGENTS.md is over its size budget, when trimming or splitting a package rules file, or when the rules file and the long-form package doc disagree. Not for writing a package README or projecting rule files into agent filenames.

Measure agent instruction files - root and per-package AGENTS.md or CLAUDE.md, rule files, skill metadata - in tokens through the provider's token-counting endpoint, report the distribution and the bytes-per-token spread, and convert existing byte budgets to token equivalents. Use when sizing or budgeting agent rule files, asking how much context instruction files cost, or setting a size limit for AGENTS.md. Not for counting tokens of a single prompt or estimating API cost of a conversation.

Verified onClaude Code

Lay out agent rules across a monorepo - a root AGENTS.md under 12 KB and 200 lines that points to path-scoped .agents/rules/*.md files, one per-package AGENTS.md of constraints only under 4 KB with the long form in .agents/PACKAGE.md under 200 lines, the CLAUDE.md and GEMINI.md links each package needs, and two checks (a context budget and no legacy .agent/ directory). Use when setting up or restructuring rules for many packages, when the root AGENTS.md has grown past its budget, or when adding budget checks for instruction files. Not for trimming one file under budget or for projecting rule files into agent filenames.

Keep one source of rules per package and project it into the filename each coding agent expects, splitting always-loaded constraints from the document read on demand. Use when a repository is read by more than one agent, or when rule files have started to disagree.

Verified onClaude Code

Run the remaining work of a long session as waves - list what is left as a numbered table with a recommended order and a mark on items that will be forgotten if deferred, run only the numbers the user picks as one wave, commit in logical units, stop at every approval gate, name what must be deployed and where, and record standing instructions such as "hold pushes" as session rules that survive context compaction. Use when the user asks what is left, says "next wave", "continue as recommended", "do 1 to 4", or when a task ends with follow-ups. Not for ranking a product backlog or planning a quarter.

Keep one living document per workstream (a shared doc, wiki page or published page) current as the work moves - update its checklist, measured results and open decisions in place, restructure sections instead of appending new text below stale text, date every measurement, and return the document's URL after each update. Use when the user asks to keep a doc, page or artifact updated, to reflect progress in the plan document, or to "update the doc" after a step. Not for writing a new one-off report or for a changelog.

Verified onClaude Code

Stop a coding session that keeps retrying the same failing approach - log every attempt with its approach and error, count attempts per approach, stop at three, run a separate diagnostic pass that reproduces the failure and lists hypotheses before any new fix, and hand only verified findings back into the retry. Use when a fix has failed two or three times, the user says the agent is looping, or the same test or error returns after repeated changes. Not for planning remaining work (followup-wave-loop) or a first failed attempt.

Verified onClaude Code

Answer "where are we" in a long agent session as a short status in four parts - done (with commit hashes and what ran, how much), in progress (subagents, builds), blocked, and the next decision the user owns - in the user's own language, with a check for work beyond the goal; a resumed or compacted session first runs a preflight (UTC time, HEAD, who owns the dirty files, running processes) and restates the goal and what remains. Use when the user asks for progress, a status, a briefing, "how far along", or when a session resumes after compaction or handover. Not for writing a release note, a postmortem or a commit message.

Verified onCodex

Split a large uniform job (dozens of files, packages or documents) across parallel subagents through one shared BRIEF file - the spec written once, chunks of 16 to 25 items with disjoint file ownership, a hard-rules block (no git, no builds or tests until one final check, touch only your own files), integration notes each agent returns, then one integration agent for the shared files and only the coordinator commits. Use when a task covers many similar items and would take hours serially, when planning parallel agents, or when the user says fan out, split across agents or batch the work. Not for one-off delegation of a single search or for running several builds at once.

Verified onClaude Code

Audit an agent graph (for example a LangGraph-style state machine) for two state bugs - a pending action forgotten after the agent asked for confirmation, and mutable objects that persist from one run into the next or into classes outside the graph - by writing the state transition table (pending, confirmed, rejected, duplicate consent, unrelated message), proposing tests that bind a consent turn to the same pending action, listing mutable objects held by state, checkpoints and outside classes and comparing references across sequential and concurrent runs, and separating intended checkpoint persistence from accidental sharing. Use when a yes to a proposed action is treated as a fresh chat, or a value changed in one run shows up in the next. Not for scoring output quality (llm-output-eval-suite) or verifying that a finished run changed the target system (agent-run-postcondition-and-replay-check).

Verified onClaude Code

Verify that an agent run really did what it reported and that a failed run can be replayed faithfully - read the target state directly after each action and compare it with the goal instead of trusting the agent's success wording, flag calls to tools that do not exist, blocked logins and failed calls counted as success, record a hash of the prompt bytes actually sent plus the code version per step, and on replay restore a checkpoint, block duplicate external side effects, mark non-reproducible tool results and report the first diverging step. Use when an agent says it finished but the result is missing, or when a failed multi-step run is replayed or traced and the cause of an output change is unclear. Not for scoring answer quality (llm-output-eval-suite), judge trust (llm-judge-calibration-and-order-check) or state bugs inside an agent graph (agent-graph-state-leak-and-confirmation-audit).

Verified onClaude Code

Design and check a benchmark protocol for model serving speed claims so the result predicts the real workload - record the real input and output length distribution, time prefill and decode separately, cross input length, shared prefix, cache state and concurrency, report per-request latency next to total throughput, account reasoning and speculative-decoding tokens against delivered tokens, count truncated and timed-out runs, and say which results come from one machine. Use when someone compares local or hosted models by a tokens-per-second figure, ranks models on short prompts, or asks whether a batch, reasoning level or speculative decoding setting helps their long-context or concurrent workload. Not for output quality evals (llm-output-eval-suite), judge trust (llm-judge-calibration-and-order-check) or running GPU workloads.

Verified onClaude Code

Decide whether an LLM judge's scores can be trusted - build a human-labelled anchor set of 40 to 60 real outputs and compare judge and human case by case, test order sensitivity by swapping candidate positions with an identical-pair repeat as the noise control, check length bias against the human choice rate, have a person verify every generated test case against a source, and record agreement and the judge prompt version with each result. Use when a judge prefers longer answers, flips when A and B are swapped, ranks change with wording, or auto-generated test cases look invented. Not for building the eval suite itself (llm-output-eval-suite), scoring one answer, or choosing a model.

Verified onClaude Code

Build and run an evaluation suite for an LLM feature - representative and adversarial cases drawn from real use, graders that judge behaviour with stated acceptance criteria, a baseline to compare against, repeated runs to see variance, and a regression gate - so prompt, model or retrieval changes are decided by numbers; also split results by task type to expose a regression hidden behind a higher average, and diagnose scores that move between identical runs (repeated trials, retrieval ids and tie order). Use when shipping or changing a prompt, model, retrieval setup or agent behaviour in a product, when quality complaints need to be measured, when a model switch raised the benchmark score or the mean pass rate but answers seem worse, or when an eval or RAG score moves between identical runs. Not for evaluating agent skills in a plugin catalog.

Verified onClaude Code

Check a comparison of quantised weights, KV-cache precision or vision-model input settings before believing a surprising score - log the settings actually applied per arm (weight quantisation, K and V cache dtype, image token budget, tile settings, thinking mode, context truncation), change one factor per run, pair results question by question with repeat runs to estimate noise, check answer extraction and truncation first, keep blank and timed-out outputs in the denominator, and report quality, memory and time together without claiming a small gap is real. Use when a lower-precision setting scores above full precision, results order implausibly, or a vision model returns blank answers after settings changed. Not for speed benchmarks (inference-benchmark-workload-protocol), slice and instability analysis (llm-output-eval-suite) or running the models.

Verified onClaude Code

Answer a browser request or handoff that another session sent in the claude-in-chrome-remote format, by running it with the browser connected to this session or replying that it cannot. Use when a message from another session starts with [claude-in-chrome-remote followed by any version such as v1 or v9.

Report whether this session can do browser work through Claude in Chrome and which route it would take, or run a small read-only smoke test of that route. Use when asked to check the claude-in-chrome-remote setup, whether a browser session is visible, or to verify the route end to end.

Choose how browser work gets done across Claude Code sessions, send it in a fixed message format and read the answer. Use when a task needs a real browser through Claude in Chrome and this session may not have one, when browser work must be handed to another session that does, or when an answer message starting [claude-in-chrome-remote v1] RES arrives.

Diagnose build or test resource contention and choose measured limits for parent fan-out, concurrent tasks and internal worker pools. Use for slow startup, loaded or shared hosts, memory budgets, or questions about how many workers to allow. Not for ordinary command syntax, selecting one test file, filtering test names or running a single test when no resource or concurrency problem is raised.

Verified onClaude CodeCodex

Design or review build-runner benchmark controls, including cold cache layers, forced execution with retained outputs, warm reuse and resolved root-dependency changes. Use when comparing build runners, auditing a cold speedup or choosing invalidation controls. Not for general code profiling, model-serving benchmarks or auditing cache implementation.

Verified onCodex

Audit build cache identity, dependency propagation, output restoration and failed-result reuse with positive and negative controls. Use when a monorepo reports suspicious cache hits, misses required rebuilds or is preparing shared cache rollout. Not for timing-only comparisons or implementing a runner migration.

Verified onCodex

Optimize a CI or pre-push gate by separating preparation, restoration, runner and task critical paths while preserving validation scope and failure behavior. Use when gates are slow, pooled checks change ordering or faster green runs execute fewer checks. Not for deleting required tests, deploying a CI service or generic application profiling.

Verified onCodex

Plan and validate a staged monorepo task-runner migration that preserves real commands, dependency ordering, selectors, caches and CI before cutover. Use when replacing a build runner, reviewing a green migration or deciding whether to keep the current runner. Not for an automatic product recommendation or a cache-only audit.

Verified onCodex

Review whether a build optimization preserves behavior across CI, process resource limits and declared workspace layouts. Use when a locally successful optimization is being reused in another monorepo or a fast scanner silently drops supported inputs. Not for a general security audit or certifying untested operating systems.

Verified onCodex

Run or review a formatter, bundler or application build-tool replacement pilot against consumer and output contracts before adoption. Use when a faster tool is proposed, normalized ASTs match or runtime tests pass but artifact parity is uncertain. Not for a site hosting migration, source-wide restyling or choosing a tool from popularity.

Verified onCodex

Separate BagIt structural completeness from checksum validity using bag inventory, manifests, fetch entries and full digest evidence. Use when fast validation passed, a transferred bag needs fixity review or missing payloads must be distinguished from changed bytes. Not for repairing, downloading or certifying a bag without evidence.

Verified onCodex

Split a large warehouse or dbt model into staging and intermediate models without changing its output, explain a model's upstream and downstream path from compiled SQL and lineage, and bring model and column descriptions in line with the real selected columns. Use when someone wants to decompose a long SQL model, onboard onto a new model's dependencies, or update schema.yml descriptions after a change. Not for proving a converted or rewritten query equals the original at scale (use sql-migration-equivalence-check) or designing a new warehouse from scratch.

Verified onClaude Code

Audit how links in a knowledge graph or a data model were created - label each link as declared, inferred or merged with its evidence, measure over-merge and under-merge on a human-labelled sample, test an inferred join for value overlap, key uniqueness, cardinality and business meaning before accepting it, and keep original mentions and a merge record so a merge can be undone with relations restored. Use when entity resolution by name or embedding similarity merged different people or types, or an LLM-suggested join between tables with no declared key multiplies rows or links wrong records. Not for deduplicating plain lists (use list-normalize-and-diff), SQL migration equivalence, or running the pipeline.

Verified onClaude Code

Fix an agent or conversation memory that returns stale facts or treats weak extractions as knowledge - record valid time and recorded time separately, link superseded facts instead of deleting them, tag every claim as hypothesis, observation or confirmed and promote it only on new evidence from a different session, answer as-of questions from the state valid at that time, and hold or drop single weak extractions. Use when a stored fact changed (a user moved, a status changed) and searches still return the old one, or repeated phrases are becoming facts. Not for entity merges or joins (use inferred-link-evidence-audit), choosing a vector store, or storing sensitive personal data.

Verified onClaude Code

Merge many lab or instrument workbooks of similar but not identical layout into one table - inventory every file and sheet (header row, columns, units, date format), group files by layout variant, write a target schema with a per-variant column mapping, convert units and dates by explicit rules, reconcile rows in against rows out per file with a stated reason for each excluded row, and spot-check cells against the source. Use when dozens or hundreds of spreadsheets from one lab need one combined table. Not for statistics and not for parsing logs or JSON (use semi-structured-field-extraction).

Verified onCodexGemini CLIGrok CLIAntigravity

Compare a saved searchable PDF's text positions with its visible scans and approved page anchors. Use when OCR words look plausible but highlights shift, rotate or omit pages, or a sidecar is offered as proof. Not for PDF reading-order audits or redaction.

Turn log lines, bench or test data and JSON documents into typed columns and a plot-ready table - inspect real samples and list every field, unit and type variation, map each source field to a column with an explicit cast and a policy for missing, string, array and wrong-type values, verify row counts, units and a hand-checked sample, and plot only from the extracted table with units on the axes. Use when someone asks to parse logs or JSON into columns, write extraction SQL or a parser, or chart measurements from raw text. Not for proving two SQL queries are equivalent (use sql-migration-equivalence-check).

Verified onClaude Code

Prove that a migrated, dialect-converted, optimized or extended SQL query or table returns the same result as the original - number the statements in source and converted script, write row-count, two-way key anti-join and null-safe per-column diff queries, seed deliberate differences to show the checker catches them, and report explained and unexplained differences apart. Use when SQL was converted between dialects, a table was replaced by a new one, a long query was rewritten for speed or got extra joins or filters, and someone asks whether it is equivalent. Not for designing a new schema or writing a query from scratch.

Verified onClaude Code

Produce the demo video a ChatGPT plugin review requires (demo_recording_url) - run every test case for real, then render the ChatGPT screens in Remotion from a measured interface reference, with the exact prompts, approval cards and answers from those runs and tool captions from the server's call log, covering every main use case and tool; verify the MP4 and host it where a reviewer can open it without signing in. Claude Motion is the alternative on a Claude Team or Enterprise seat. Use when the user needs a demo recording, walkthrough or intro video for an OpenAI or Claude directory submission, or asks how to make one with Remotion or Claude Motion. Not for marketing videos that show results the plugin does not produce.

Verified onClaude Code

Write one listing source and fit it to both directory forms - OpenAI interface fields (display name and short description 30 characters, long description 4000, up to 3 default prompts, category, four policy URLs) and the Claude connector listing (name 100, one-liner 200, description 2000, 1 to 5 categories, use cases, data handling answers) - with counted lengths and the naming rules reviewers reject on. Use when the user writes or fixes a plugin or connector listing, description, starter prompts or data handling answers for the ChatGPT or Claude directory. Not for website marketing copy.

Verified onClaude Code

Write the review material directory reviewers test with - exactly five positive and three negative test cases in the OpenAI review shape (prompt, tools_triggered, expected_behavior), the reviewer account sheet kept out of the package, and the matching Claude Test & launch instructions - all from the same source. Use when a ChatGPT plugin or Claude connector submission needs test cases, negative cases, reviewer credentials or reviewer instructions. Not for writing the plugin's own eval suite.

Verified onClaude Code

Turn one plugin folder (skills, manifest, optional remote MCP server) into what the ChatGPT plugin directory and the Claude directory each accept - the OpenAI ZIP with plugin.json, mcp.json, assets and review fields, the Claude plugin folder on GitHub, and the Claude connector submission when there is a server - and check every limit before upload. Use when the user wants to submit or update a plugin, skill pack or MCP server in the OpenAI or Claude directory, or asks what goes in the ZIP. Not for writing the listing text or test cases themselves.

Verified onClaude Code

Judge whether a lifecycle hook actually works in a given coding agent, and what to do when two agents disagree. Use when writing hooks meant to run in more than one agent, or when a hook appears to do nothing.

Verified onClaude Code

Diagnose MCP client behavior from recorded initialization, paginated tool discovery, client exposure and actual harmless calls. Use when a connected server exposes no tools, listed tools fail when called, clients differ, or headless permission behavior needs measurement. Not for designing tool schemas, implementing a server, or changing live authentication.

Verified onCodex

Check that a 3D explorer, configurator or guided tour matches its underlying data - define the schema of the selectable state, round-trip a selection to the order or data record, recompute price and check disallowed combinations and upload or input limits against the schema, verify that guide text and the highlighted object refer to the same data item, mark scale distortions, confirm user data stays local when claimed, and require a non-3D fallback. Use when a builder shows a configurator order, selection state or explorer data and asks whether the 3D view, the order and the numbers agree. Not for physics accuracy or frame time (interactive-simulation-reference-check) or for building the viewer.

Verified onClaude Code

Make an AI-built interactive demo run outside the assistant preview or artifact sandbox - fix the delivery target (opened HTML file, offline bundle or static hosting), inventory every dependency, host-provided API, asset and storage call, replace each environment-specific one, and test in a clean browser with the exact environments reported. Use when a demo works in a chat preview but fails when copied out, or when a creator wants a standalone file to share. Not for checking simulation accuracy, validating configurator data, deploying a production service or a launch audit.

Verified onCodex

Validate a browser simulation or visual demo (physics, flocking, cloth, terrain, shader, data sonification) against a known result and measured performance - compute the error against an analytic or reference value, test timestep convergence, check units between graph and 3D view, rule limits, state persistence and reset, data-to-view mapping, and frame time and memory on stated devices, and never report a figure that was not measured. Use when someone built an interactive demo with an AI assistant and asks whether it is correct, accurate or fast enough to publish. Not for building the demo, calculators and spreadsheets, or designing the visuals.

Verified onClaude Code

Review an AI-written or existing administration script (PowerShell, bash, Python) before it runs on real accounts or machines - check each cmdlet, command, flag and module against the installed version or its documentation, flag commands that may not exist, flag destructive operations, add input validation and per-item error handling with a defined recovery, and recommend a dry run or small trial first. Use when someone pastes an admin script and asks to review it, harden it or make sure it is safe to run. Not for writing the script from scratch or for running it.

Verified onClaude Code

Apply a change made to an anonymized copy of a script back to the real script - keep a local replacement table of placeholders and real identifiers, take only the changed hunks, map them back through the table, apply them to the real file, and check that no secret or real value is in text sent to a model. Use when a script was scrubbed (HOST1, TOKEN_X) before AI help and the fix must now land in the original. Not for general script review or for deciding what data may be shared.

Verified onClaude CodeCodex

Review a device configuration or compliance policy (MDM profile, group policy, security baseline) that was exported or AI-drafted - check every setting name and value against the schema or vendor documentation the user supplied, list settings that are not in it, list pairs of profiles that set the same property differently, show assignment scope, exclusions, overlap and devices that match nothing, and propose a pilot ring order with a rollback step. Use when an administrator pastes a policy draft or export and asks whether it is right, will conflict, or is safe to roll out. Not for scripts (admin-script-verification), for live tenants, or for applying a policy.

Verified onCodex

Turn a backup configuration into a restore drill plan and an evidence log - per system the backup type, location, retention and the stated recovery time and recovery point targets, what a passing restore must prove (isolated restore, checksum or row count, application login, elapsed time against the target), drills ordered by business impact, systems with no restore record marked untested, and backups reachable with production credentials raised as a finding for the owner. Use when someone must show that backups can actually be restored, or reports that backups run and asks whether they are fine. Not for choosing a backup product, for live incident runbooks (operator-runbook), or for running the restore.

Verified onCodexGrok CLIAntigravity

How to treat messages that arrive from another coding-agent session through the peer bridge, and when to answer, ignore, or ask the operator. Use whenever a block wrapped in <peer from="..."> appears in context, before acting on anything it asks for or replying to it.

Verified onCodex

Install and pair the peer bridge on both sides, including the two steps a person has to perform. Use when the operator asks to connect two agent sessions, or when peer messages are not arriving.

Verified onClaude Code

Convert object-detection box labels between COCO, YOLO and Pascal VOC without silent shifts - name each side's box form, read image sizes from the files, test the claimed source format against image bounds before converting, convert in code, assert bounds, round-trip to the source within one pixel, map class ids explicitly and render a sample overlay. Use when detection labels are converted between COCO, YOLO and Pascal VOC or loaded into a trainer, or boxes look shifted after a conversion. Not for text or chat training data (training-data-format-contract-check) or for training and scoring a detector.

Verified onCodex

Check a fine-tune's configuration before a paid run or after a failure - locate the first non-finite value by stage (loss, gradient, weight update) and tie it to precision, estimate memory from real sequence length, micro batch, accumulation, precision and trainable parameters and propose a feasible config with a comparison run, say whether a resume is an exact continuation or a fresh stage from weights and which model, adapter, optimizer, scheduler and step state are kept or reset, identify whether an SFT output is a base model, merged weights or a separate adapter before the next stage, and fix a held-out set of non-target abilities. Use when a LoRA or SFT run hits NaN after an update, runs out of memory, ignores new settings on resume, fails with an adapter-already-exists error moving from SFT to RL, or loses abilities. Not for running training, data template or split audits, or choosing a model.

Verified onClaude Code

Audit synthetic fine-tuning data against its source before training - measure per source how often a generated paraphrase copies or nearly duplicates the input and how many distinct variants survive, link each synthetic answer to a source fact and list facts omitted, added or repeated, validate every tool name and argument in tool-call targets against the real catalog and the task intent including calls that should not be made, and build unanswerable question and unrelated-context pairs with a balance ratio. Use when a model trained on back-translated or generated data copies its input, recalls nothing from synthetic QA, calls invented tools or arguments, or never says it does not know. Not for running training, split leakage, template or mask checks, or choosing hyperparameters.

Verified onClaude Code

Check and convert fine-tuning data against the target trainer or provider schema before any upload or training run - name the schema first, validate role, field name and content type per row, serialise code, newlines and non-ASCII text as one JSON record, reload with the real loader and compare row counts, check tool-call turns and ids, and report rejected rows by line number. Use when a fine-tuning JSONL is refused at upload, loads with fake or corrupted rows, or tool-calling chat data raises a type error in the chat template. Not for running the training job, choosing hyperparameters or evaluating the trained model.

Verified onClaude Code

Audit a classifier's or fine-tune's data when the score looks too high, plateaus low, or the errors look like wrong labels - trace every sample to its source group (patient, slide, document, author), split by that group and list groups present in more than one split, review the model's mistakes against the label definition with a correction ledger (original label, proposed label, reason, approver), count examples per class with source diversity and length before touching hyperparameters, and state which later data was reused for tuning so the final score is not called independent. Use when a random split may mix one source across train and validation, labels may be wrong, or classes are tiny or uneven. Not for running training, template or tokenization checks, upload format errors, or clinical or legal sign-off.

Verified onClaude Code

Audit fine-tuning data as the trainer will really see it, before an expensive SFT run - render sample rows with the exact saved chat template, show string, token ids and label mask side by side, count per-row length against the context limit with cut position and remaining supervised tokens, flag rows with zero targets, check packing for cross-sample attention, loss mask and position reset, and compare the training render with what the serving template emits. Use when valid JSONL may still change under the template, truncation or packing, a template strips text, or training and serving prompts may differ. Not for running training, schema or upload errors, split leakage, or choosing hyperparameters.

Verified onClaude Code

Check a personal-data scrub of a text corpus before it goes into training or is shared - list the PII types in scope, require pattern detection plus entity detection, measure recall per type on a hand-labelled sample from the counts supplied, choose a replacement strategy (consistent pseudonyms or type tags), keep the re-identification map outside the dataset in an access-controlled place, and report residual risk. Use when someone says a corpus was cleaned with regular expressions, or asks whether text is safe to train on or hand to a partner. Not for legal advice on lawful basis or consent (pipa-review), for splitting or label audits, or for running the scrub; the privacy officer decides release.

Verified onCodex

Set up and read the per-package check that runs after an edit - choose the marker file and command for the stack, confirm the package really has that check, tune the debounce, test the four payloads, read the trust risk of the repo config, and when a failure is reported fix the named package instead of running the whole-repo check. Use when configuring monorepo-package-validate, when a package check failure appears after an edit, when the check is too noisy or never speaks, or for a Cargo, Python or Go workspace. Not for designing CI pipelines or for fixing the root workspace build.

Verified onCodex

Review a growing skill catalog before payload limits force a redesign - measure list, body and bundle bytes separately, plan pagination and lazy reads, bind caches to revisions and authorization, and stage a migration with rollback and boundary checks. Use when a catalog approaches a response size cap, many skills exhaust context, or a team asks how to scale catalog synchronization. Not for writing skills or deploying an HTTP skills server.

Verified onClaude Code

Handle a data quality incident in a collection or analytics pipeline - detect missing, late or duplicated data against expected volumes, bound the affected window and tables, find the cause stage by stage, stop further damage, plan and run an idempotent backfill, verify the counts after, and tell the people who used the bad numbers. Use when a dashboard drops to zero, numbers look wrong, a crawl or ingestion job silently stopped, or data arrived twice. Not for designing a new pipeline.

Verified onClaude Code

Accept changed MCP tools through an outside agent - turn the changed tool list into a request document with scenarios, acceptance criteria and a measurement script that an external agent or customer model runs, take back its report as pass, fail or not run per scenario, separate real defects from client limits and unclear requests, fix, bump the version and deploy, then reissue the request for the failed and not-run scenarios. Use after changing or adding MCP tools, before telling users a tool works, or when asked for a pilot, acceptance test or request document for an external agent. Not for proving a customer server's isolation and protections, and not for designing tool schemas.

Verified onClaude Code

Run several agents implementing in one shared working tree - split the work into lanes, give shared files (router, entry point, provisioning SQL, required-migration pin, copy catalog, contracts index) to one integration agent while lanes write the lines they need into integration notes, fix migration timestamp ranges per lane, serialise heavy checks with a host lock, and after integration reconcile the migration pin, clause numbers and extended enum or check constraints. Use when planning or running parallel agents on one checkout, or when merging their lanes. Not for a single agent working alone or for separate repositories.

Verified onClaude Code

Stand up one HTTP skills server per customer with pmcp's createSkillHttpHandler and prove isolation and protections before a pilot opens - own handler, port and marketplace per customer, no default roots, bearer auth, allowedHosts, body cap and rate limit - then run request checks (401s, catalog isolation, unknown_skill, redacted .mcp.json, 403 origin and host, 413, 429) and drive a real client headless through the customer's business cases, reporting N/N with the case list. Use before opening a pilot or a customer's skills server, after changing a customer server's configuration, or when asked to prove one customer cannot see another's skills. Not for designing the skills themselves or for a single-tenant staff server.

Verified onClaude Code

Host a company skill catalog over HTTP so non-developers reach it from a hosted chat app through a custom MCP connector - wrap pmcp's createSkillHttpHandler with authorize, tiersFor and a private cacheScope, deploy it as a container or a worker, and write the data-flow disclosure a security review asks for. Use when staff without a repository should use the company's skills, when adding a skills server as a custom connector, or when a skills server needs security approval. Not for a local stdio setup in a repository or for designing a general MCP server.

Verified onClaude Code

Threat-model a system with STRIDE - draw the data flow with trust boundaries, list threats per element (spoofing, tampering, repudiation, information disclosure, denial of service, elevation of privilege), rate each by likelihood and impact, map existing controls, and leave a ranked list of gaps with owners and tests. Use when designing or reviewing an auth service, an API, a payment flow, a browser extension or anything that crosses a trust boundary. Not for scanning code for known vulnerabilities.

Verified onClaude Code

Move a repository's scattered agent files - CLAUDE.md, .cursorrules, .windsurfrules, copilot-instructions.md, .claude/skills, .claude/agents, .claude/commands, .gemini/agents and a root .agent/ folder - into one AGENTS.md, .agents/ and pmcp.toml without losing a line, keeping historical citations as they are, and finishing with a clean pmcp project --check. Use when consolidating agent instructions from several tools into one source, when adopting pmcp in a repository that already has agent files, or when agent files have drifted apart. Not for a new empty repository or for trimming an oversized AGENTS.md.

Write a CI job that runs Claude Code, Gemini CLI, Codex, Grok Build and Antigravity headless against a project MCP server, giving each the permission setting it actually honours, asserting on the tool call rather than the exit code, and gating with pmcp project --check and pmcp doctor. Use when adding agent CLIs to CI, when a headless agent run skips an MCP tool call, or when a permission flag seems to do nothing. Not for interactive sessions or for writing the MCP server itself.

Verified onClaude Code

Package skills, commands, subagents, hooks and MCP servers as a plugin in a plugin marketplace laid out the way pmcp's catalog reads it - marketplace.json with a tier, plugin.json, skills/, commands/, agents/, hooks/hooks.json and .mcp.json - and confirm every asset is listed with pmcp list and pmcp validate. Use when turning a folder of skills into a marketplace plugin, when a plugin's assets do not show up in a skill catalog, or when setting the tier of a marketplace. Not for writing a single skill's content or for hosting the catalog over HTTP.

Verified onClaude Code

Set up a repository so every coding agent reads the same instructions, skills, subagents, rules and MCP servers - one AGENTS.md with CLAUDE.md and GEMINI.md linked to it, sources under .agents/, a pmcp.toml with [targets] and an [mcp.*] entry, then pmcp project and pmcp doctor, plus the one-time trust step each agent CLI needs. Use when starting a new repository the pmcp way, adding pmcp to a repository that has no agent assets yet, or when asked to write a pmcp.toml. Not for moving existing scattered agent files (migrate them instead) or for hosting a skills server.

Find, check and load the right skill from a skill catalog server instead of guessing. Use when a task might have a dedicated skill, when the pmcp tools (skill_find, skill_catalog, skill_describe, skill_call, skill_read) are available, or when a search for a skill came back empty.

Verified onClaude Code

Lay out one new skill folder that passes the Agent Skills specification - SKILL.md frontmatter within the limits pmcp validate enforces, references/, scripts/ and assets/ only where needed, and evals/ with a normal, an exception and a missing-information case - then gate it with pmcp validate. Use when creating the folder for a new skill, when pmcp validate rejects a skill, or when a skill needs its first eval cases. Not for deciding what the skill should teach or for packaging several skills as a plugin.

Run terms of service, privacy policies and consent documents as versioned files - one markdown file per version and language with frontmatter for version, effective date and kind of change - so the version in force is chosen by date, a fix that does not change what users agreed to keeps their consent, translations never pose as the binding text, and public pages show every version and what each one changed. Use when a policy edit forced every user to consent again, when designing how legal documents are stored and gated at sign-in, when users need past versions and a change history, or when adding a translation or announcing a revision ahead of its effective date. Not for drafting the legal wording (use terms-privacy-drafting) or comparing two contract drafts for a reviewer (use document-version-change-impact).

Verified onCodex

Keep one registry of the outside services a product sends data to - who they are, the code signals that show a service is in use, whether personal data reaches them, live or test, and which policy documents must name them - and a blocking check that fails when code, registry and the documents in force drift apart. Use when a privacy policy's processor or overseas-transfer list has to stay true to the code, when an audit found a vendor no document mentions, when a test-key vendor is about to go live, or when asked how to stop undisclosed vendors from shipping. Not for drafting the policy text (use terms-privacy-drafting) or for a one-off connector disclosure for a security review (use data-flow-disclosure-pack).

Verified onCodex

Judge whether a written agent rule will actually hold, and decide whether to enforce it, keep it as guidance, or delete it. Use when adding a rule to an agent instruction file, or when the same mistake keeps recurring despite being documented.

Verified onClaude Code

Hand a bounded implementation or audit job to a different agent CLI using a brief file and a result file, explicit permissions, a single launch and enforced limits. Use when asked to offload work to Claude Code, Codex CLI, Gemini CLI, Grok CLI or Antigravity CLI, or prepare its runnable handoff. For an independent second opinion on an existing result or release claim, use cross-runtime-second-opinion. Verify returned artifacts before relying on completion claims.

Verified onCodex

Obtain or prepare an independent, adversarial second opinion from another agent runtime before trusting a result or approving a release. Use for peer review, requests for another CLI to challenge an existing result, or deciding whether model agreement is sufficient evidence, including reviews whose execution is deferred. Provide a runnable file handoff and a separate local acceptance probe. General work offloading belongs to cross-runtime-delegation.

Verified onCodexGrok CLI

Route live research through another agent runtime when the current one cannot search. Use when search is unavailable or exhausted, or a peer offers a research summary. Send a bounded file brief, demand a full report with claim-level citations, read the report and audit supporting pages; mark unsupported facts UNCONFIRMED instead of promoting a peer summary into evidence.

Verified onCodex

Compare subscription allowance and token billing for the same workload before routing it to an agent runtime. Use when choosing between an included CLI allowance and a paid API route, estimating a peer job's cost, or evaluating a runtime subscription. Use real input, cache, output and tool counts, preserve the quota floor, and enforce a spend cap before paid execution.

Verified onCodex

Check remaining allowance and reset times before choosing an installed agent runtime for a job. Use when a runtime is near its quota, several CLIs are available, or asked where a job can run without exhausting a subscription. Normalize every applicable window, reject stale or unknown readings, reserve a floor after the job, and save the routing evidence.

Verified onCodexGrok CLI

Respond to a secret-exfil guard refusal and tune its configuration - answer the user's question without printing credential values (existence, line count, key names, digests), hand them the command to run themselves, and extend secret-exfil-guard.json with project paths and value-printing CLIs, then prove one allowed and one refused command. Use when a command that reads or sends .env, credentials, key files or the whole environment was refused, when asked to cat a secret or upload a credentials file, or when adding a project path or CLI to the guard. Not for rotating a leaked secret or for general security review.

Review archive member, link and destination policies for a named extractor before extraction. Use when assessing an untrusted TAR or ZIP handoff, a filter refusal or whether listed names prove containment. Not for executing suspicious archives or certifying universal extraction safety.

Verified onCodex

Audit a pasted firewall or ACL rule export against the approved standard the user supplies, finding standard violations, typos, and rules that are shadowed or unreachable because of rule order, with the real allowed path stated per flagged rule. Use when an engineer pastes an ordered rule list plus a written standard and wants violations and order problems found. Not for scanning or probing live networks or third parties, writing new rules, approving a change, or log analysis (see security-log-investigation).

Verified onClaude Code

Decode and explain a quarantined obfuscated script the user pasted by static reading only, layer by layer, with the transform recorded per layer, what each layer does, indicators found and not recoverable, and defanged URLs and hosts. Use when a responder has a suspicious script sample and wants to understand what it does for their incident process. Not for running, detonating or testing a sample, writing or improving malware, evasion or obfuscation, producing working exploit code, attribution of an actor, or verdicts that a host is clean.

Verified onClaude Code

Review applied PDF redaction and residual-content evidence in the final file, separating visible coverage from deletion and hidden-data checks. Use when checking a redacted PDF before sharing or assessing negative search results. Not for forensic recovery, reidentification or a guarantee of irrecoverability.

Verified onCodex

Audit what a repository, npm package or docs site will expose before it goes public - a census of internal paths, hostnames, private identifiers, non-English or internal comments, and auth or billing internals across published tarballs, READMEs, sourcemap sourcesContent and docs, plus the license, commit author identity and repository name - with the deny list of internal terms kept private and outside the published tree. Use before making a repository public, publishing a package, mirroring source to a public repo, or opening a docs site. Not for secret scanning of a private repo's history alone or for a general security review.

Verified onClaude Code

Draft customer security questionnaire answers or a periodic access-review packet only from approved policy documents and ledger records the user supplies, linking each answer or finding to a record, marking unanswered items and stale documents, and requiring approver sign-off. Use when a team answers a customer security questionnaire or prepares a quarterly access-review packet from approved records. Not for inventing controls, certifying compliance, legal or contractual commitments, audit opinions, or incident analysis (see security-log-investigation).

Verified onClaude Code

Turn pasted logs, an unfamiliar log format, an investigation question or a packet-capture summary into a sourced timeline, a field-to-purpose map or a SIEM search, with every link marked observed, inferred or unknown. Use when an analyst needs an incident timeline from raw log lines, wants to know what a new log's fields can detect, asks for a search query from a question, or wants capture retransmits and errors tied to hypotheses. Not for security advice, declaring an incident confirmed or closed, containment or remediation decisions, or running queries against live systems.

Verified onClaude Code

Match a pasted vulnerability advisory or threat notice against the user's asset or dependency inventory and, for dependencies, the code that calls them, producing exposed, not exposed and unknown lists with the quoted trigger condition and evidence. Use when a security owner asks which of their assets or applications an advisory affects, or whether a vulnerable dependency is actually reachable. Not for scanning systems, exploiting anything, declaring an asset safe, patch scheduling decisions, or log analysis (see security-log-investigation).

Verified onClaude Code

Evaluate a plugin's skill activation and answer quality with Antigravity CLI, paired with and without the plugin, scored cases, JSON results, and an HTML report. Use for plugin regressions and release checks.

Evaluate plugin skill activation and answer quality with Claude Code native plugin eval, paired baseline runs, case graders, common JSON results, and local HTML reports. Use when a user requests Claude plugin regression or release evaluation.

Evaluate a plugin's skill activation and answer quality with Codex CLI, paired with and without the plugin, scored cases, JSON results, and an HTML report. Use for plugin regressions and release checks.

Write a new agent skill (SKILL.md) that changes what the model does - eval cases first, a description that triggers on the right requests and not on near misses, a body limited to what the model gets wrong without it, and frontmatter that passes the marketplace schema. Use when the user asks to create, improve or review a skill, a SKILL.md or a plugin skill. Not for writing general documentation or system prompts.

Verified onClaude CodeCodex

Read a skill eval result table and decide what to change next - description first when the skill never fired, re-measure both arms when with-only graders inflate scores, rerun before editing on a single judge failure, keep worked examples independent of the eval prompt, and stop after two fix rounds. Use when a skill's eval scores are given, when a score is 0 with the skill, when without and with are both 1.00, or when a worked example may mirror the eval case. Not for writing a skill from scratch or for installing a plugin.

Verified onCodex

Evaluate a plugin's skill activation and answer quality with Gemini CLI, paired with and without the plugin, scored cases, JSON results, and an HTML report. Use for plugin regressions and release checks.

Evaluate a plugin's skill activation and answer quality with Grok CLI, paired with and without the plugin, scored cases, JSON results, and an HTML report. Use for plugin regressions and release checks.

Set up how an agency or consultant delivers agent skills built for its own clients - one bundle per client with pinned versions, delivered as a private npm package or a per-client MCP endpoint whose skills a host can copy with skill_bundle, and strict isolation so no file, reference or catalog is shared across clients. Use when building skills for several client companies, when packaging client-specific skills for handover, or when a client asks how skills reach their staff. Not for an internal company skill repository or for writing the skills themselves.

Verified onClaude Code

Set up a company skill repository where approval runs on Git - main holds only approved skills, edits happen on draft branches, CODEOWNERS names the approvers, releases are vN tags with a CHANGELOG, CI runs pmcp validate and the eval cases, teams pin an approved version, and rollback re-projects from the previous tag. Use when a team wants an approval process for shared agent skills, when skills are copied around with no record of who approved what, or when teams need to pin and roll back skill versions. Not for writing one skill's content or for measuring a skill's impact.

Verified onClaude Code

Decide whether a skill or plugin is ready to list - run its eval suite with and without it, read the per-case deltas and invocation results, separate grader mistakes from skill failures, and apply a release gate with a cost cap. Use when the user asks to evaluate, benchmark, gate or release a skill or plugin, or to interpret a claude plugin eval report. Not for writing the skill itself.

Verified onClaude Code

Measure whether a business skill actually helps a team - build normal, exception and missing-information test cases from real work, run the same inputs before and after adopting the skill, and record processing time, re-edits, omissions, wrong judgements and human interventions per case, then report the change case by case with its limits. Use when the user asks whether a skill or AI workflow saved time or reduced errors, wants a before and after comparison, or needs numbers to justify rolling a skill out. Not for model benchmark suites or for interpreting a plugin eval report.

Verified onClaude Code

Update a business skill when the rule behind it changes - edit the skill and its references to the new rule version, add an eval case for the change, re-run its evals, find and mark the old copies and saved prompts people still use, and write a change note for the team with what changed, the effective date and what to stop using. Use when a policy, price, form or procedure changed and a skill, prompt or template built on the old one must follow, or when people get different answers after a rule change. Not for writing a new skill or for gating a first release.

Verified onClaude Code

Turn a procedure that lives in one experienced person's head into an agent skill - interview the expert with cases, extract the rules, exceptions, inputs, outputs and stop conditions, confirm them back, then write SKILL.md plus references with a worked example set (normal, exception and missing-information cases). Use when the user wants to capture how a senior colleague handles a task, document tacit know-how as a skill, or prepare an interview to build a skill from an expert. Not for writing a skill from an existing written manual or for general meeting notes.

Verified onClaude Code

Verify each finding of an AI or human code review against the diff and the code that would run, mark it confirmed, refuted or unverified with the call path or test that settles it, and list the docs, config keys and examples the diff made wrong with both lines quoted. Use when a review lists claimed bugs that must be checked before acting, when a review sounds plausible but cites code you cannot find, or when docs must be checked after a behaviour change. Not for writing a first review, style feedback, or mapping changes to test cases.

Verified onClaude Code

Decide whether a failing regression test is flaky or a product defect, and write the investigation note with confirmed facts apart from hypotheses - separate the symptom, the required behavior and the current behavior, collect rerun evidence, and label a cause as a hypothesis until it is reproduced. Use when a test fails intermittently, a pipeline fails and someone asks flaky or real bug, or a defect investigation note must be written. Not for writing new test cases, fixing the code, or incident postmortems for an outage.

Verified onClaude Code

Resolve a git merge or rebase conflict so both sides' intent survives - state each side's intent in one sentence from its commit message and diff, write a resolution that keeps both, list the cases where both cannot hold instead of choosing silently, confirm no conflict markers remain and the names used on both sides still exist, and name a test with a numeric example per intent. Use when a developer asks to resolve a conflict, or to check a resolution already made. Not for choosing a branching strategy, rewriting history, or reviewing unrelated code.

Verified onClaude Code

Triage a maintainer's backlog with evidence - decide whether an old issue is still open by checking linked pull requests, later commits and a reproduction on the current version, propose a duplicate only when symptom, trigger and root cause match, and review a pull request against the issue it claims to fix line by line, marking each verdict verified or unverified. Use when a maintainer works through stale issues, possible duplicate reports, or a pull request that says it fixes an issue. Not for a tester judging a flaky test, writing the fix, or deciding to close, link or merge for the maintainer.

Verified onClaude Code

Run or judge a profile, change and benchmark loop - name the bottleneck from a profile with numbers, change one thing per experiment against a stored baseline in the same environment, require the tests to pass and several repeats, report variance and regressions elsewhere, and reject a gain that sits inside the noise. Use when someone claims a speedup from one run, asks to optimize slow code, or wants a benchmark result checked before it is accepted. Not for load testing a service, capacity planning or choosing a profiler product.

Verified onClaude Code

Verify that CLI onboarding commands identify the available package, version and executable without workspace or PATH masking. Use when reviewing a public CLI quickstart, a scoped package has a different binary name, or local success may not reproduce for users. Not for publishing releases, upgrading dependencies, selecting application starters or changing authentication.

Verified onCodex

Update test cases or a test strategy after a story, design or code change - build a requirement to change to case trace table, list changed behaviors with no case and cases with no behavior, flag duplicates, and propose manual versus automated per case for QA approval. Use when a story, diff or design changes and the existing cases must be revised or reviewed for gaps and redundancy. Not for writing test code or choosing a test framework.

Verified onClaude Code

Check whether a reused project starter or boilerplate is still sound - list the dependencies each feature adds so unneeded ones can be cut, compare declared versions and the lockfile with the target runtime and the current support policy fetched at run time, generate in a clean directory then install, build and run a minimal check, report the first failure, and record the result with its date. Use when a template is copied for every project, is years old, or may be bloated or no longer install. Not for writing repository agent rules, picking a framework, or upgrading an application in production.

Verified onClaude Code

Build test support that stays honest - an API mock written from the API definition with its error statuses, pagination and latency, a fixture that takes and releases a port and proves recovery, and a test-runtime speedup that keeps every assertion and reports time before and after. Use when a developer asks for a realistic mock, a port-occupying or resource-holding fixture, or faster tests that reload large data. Not for deciding which cases to write, choosing a framework, or load testing a live service.

Verified onClaude Code

Work safely in a repository other agent sessions are also working in, including how to commit without capturing someone else's changes. Use when other working trees exist, when the tree is already dirty, or when a gate fails on files you did not touch.

Verified onClaude Code