Claude Skills
The open format is called Agent Skills and works in Claude Code, Codex, Cursor and other agents — most people know it as Claude Skills.
Every Agent Skill we could find on GitHub, deduplicated by content. 80 149 files from 1 774 authors, of which 62 489 are unique — the rest is the same skill repackaged into someone else's repository. For each one: what it weighs in tokens, whether it ships runnable scripts, and which MCP servers it needs.
> tutorials, how-tos, reference, explanation, and developer docs (README, CONTRIBUTING, ADRs). Detects the docs stack; gates on the outline before writing prose; verifies every claim against the code before it ships. Triggers on "absolute docs", "write docs", "write a tutorial", "write a README", "document this", "improve this doc", "audit our docs".
> Use when the user wants to simplify, clean up, refactor, tidy, or refine code — their staged/unstaged git changes or a target file/path. Reduces complexity, flattens nesting, removes redundancy and dead code, scores each change by value "simplify", "simplify this", "simplify my code/changes", "clean up", "clean this up", "clean up my changes", "refactor this", "make this cleaner", "tidy this up", "reduce complexity", "flatten this", "remove dead code", "make it more readable", "polish before commit", or "absolute simplify". Acts on your working diff; for repo-wide dead code use absolute-prune; for lint/type debt use absolute-debt.
> clarify pass (3–5 questions, not a grill) → reviewed design doc written to docs/plans/ → independent scored review → stop. No task board, no build. Use when you want a spec to discuss, hand off, or review before committing to implementation. Chains into absolute-work when ready to build. Triggers on "absolute spec", "write a spec", "spec out this feature", "draft a design doc", "I want a spec to hand off / review, don't build it yet".
> reviewed spec → dependency-graphed task board → safe-wave TDD execution → verification → converge. Handles features, bugs, refactors, greenfield projects, planning breakdowns, and migrations. Triggers on "absolute work", "build this end-to-end", "plan and build", "break this into tasks", "grill me on this plan", "pick up this ticket", "run this migration".
> Build polished, intentional UIs with concrete CSS/Tailwind values — typography, color, layout, spacing, dark mode, accessibility, animations, components. Encodes specific, opinionated rules with exact values, not vague advice. Covers buttons, cards, forms, tables, navigation, dashboards, landing pages, onboarding, and more. Triggers on "absolute ui", "design a UI", "style this component", "make this less like AI slop", "fix the spacing/typography/color", "dark mode", "design a pricing page", "make this look better".
> Triggers on "absolute upgrade", "upgrade our dependencies", "bump deps", "update packages", "move off the deprecated X", "clear the Dependabot backlog".
Best practices for integrating TanStack Query with TanStack Router and TanStack Start. Patterns for full-stack data flow, SSR, and caching coordination.
TanStack Query (React Query) best practices for data fetching, caching, mutations, and server state management. Activate when building data-driven React applications with server state.
TanStack Router best practices for type-safe routing, data loading, search params, and navigation. Activate when building React applications with complex routing needs.
TanStack Start best practices for full-stack React applications. Server functions, middleware, SSR, authentication, and deployment patterns. Activate when building full-stack apps with TanStack Start.
> Ultra-compressed communication mode. Cuts token usage ~75% by speaking like caveman wenyan-lite, wenyan-full, wenyan-ultra. Use when user says "caveman mode", "talk like caveman", "use caveman", "less tokens", "be brief", or invokes /caveman. Also auto-triggers when token efficiency is requested.
> Compress natural language memory files (CLAUDE.md, todos, preferences) into caveman format to save input tokens. Preserves all technical substance, code, URLs, and structure. Compressed version overwrites the original file. Human-readable backup saved as FILE.original.md.
> Ultra-compressed commit message generator. Cuts noise from commit messages while preserving intent and reasoning. Conventional Commits format. Subject ≤50 chars, body only when "why" isn't obvious. Use when user says "write a commit", "commit message", "generate commit", "/commit", or invokes /caveman-commit. Auto-triggers when staging changes.
> Quick-reference card for all caveman modes, skills, and commands. "caveman help", "what caveman commands", "how do I use caveman".
Add or resolve an `availability` frontmatter field for a docs.getdbt.com page, so the page renders the correct applicability badge (where a feature runs, what plan/access it needs, which dbt engine version). Use when asked to add an availability badge, mark a page as platform/local/Enterprise-only, or figure out what availability value a feature needs.
Use when converting a Notion markdown export to a Docusaurus blog post for the dbt docs site, or when given a path to a Notion export folder/file.
Scaffold a new skeleton markdown doc for docs.getdbt.com — standard guide, reference page, feature/product page, or release note entry. Use when asked to create a new doc page, scaffold a doc, start a new page, or add a page to the sidebar.
| Use this skill to triage, score, and label GitHub issues in the docs.getdbt.com repo. Triggers when the user asks to triage issues, rank issues by priority or size, label GitHub issues, review the issue backlog, or score open docs issues. Also use this skill when the user says things like "what should we work on next", "rank our issues", "apply priority labels", "apply size labels", or "what are our highest priority issues". Always use this skill for any request involving prioritizing or sizing docs GitHub issues.
| Use this skill to triage, score, and label GitHub issues in the docs.getdbt.com repo. Triggers when the user asks to triage issues, rank issues by priority or size, label GitHub issues, review the issue backlog, or score open docs issues. Also use this skill when the user says things like "what should we work on next", "rank our issues", "apply priority labels", "apply size labels", or "what are our highest priority issues". Always use this skill for any request involving prioritizing or sizing docs GitHub issues.
Use when drafting or revising internal technical design docs, review notes, architecture writeups, or sharing posts in Chinese, especially when the tone should stay evidence-backed, calm, and suitable for peers or leaders rather than marketing or literary writing.
>- Analyze an agent transcript for generalizable learnings about next-browser CLI usage, then propose SKILL.md updates, feature requests, and bug fixes.
Weekly, on a Friday, read only everywhere. Scores the week from the daily metrics ledger alone, with the screen and the date range beside every number, then writes the ranked paste ready list of what to change in the account next week. Each line names one change, the exact screen, the current value, the proposed value, and the evidence row it came from. It proposes no change it cannot source, it names no action it took because it takes none, and it spends only where you released it.
Weekdays. The only routine in this kit that opens a screen inside an account that can spend, and everything it does there is read. It confirms the primary conversion event still fired inside the read window before it trusts any other figure, then reads yesterday's spend, delivery, and result rows at account, campaign, ad set, and creative level and appends one row per object per day to the daily metrics ledger with the screen, the date range, and the read date beside every figure. It creates nothing, saves nothing, activates nothing, and spends nothing, unless you released the channel.
Weekdays. Works the next ready card on the board, one card per run, and assembles what it needs as a complete local file ready to paste: a campaign build sheet, a negative keyword file, a conversion action specification, an audience definition, or an upload packet. The whole step happens in a text editor until you release the ad account in RELEASES.md, and then Step 6a publishes the approved package through a connection you already have and writes a receipt for every id the platform returns.
Weekdays. Produces one creative set per run against evidence rather than taste, reading the standing doctrine for the angles currently earning and the metrics ledger for what has decayed. It writes the images and the exact strings into a dated local set folder with the destination screen named, files one card, and uploads nothing to any account. It never opens an account screen, sends only where you released the channel, spends only where you released it, and never touches a credential.
Weekdays, file work only, no browser at all. Reads every run record, folds the metrics, change, and creative ledgers, turns yesterday's ticks into facts a machine can count, folds the card inbox, rewrites the board, and writes the short morning brief the member opens first. It holds every outbound action unless you released the channel, and it never touches a credential.
Runs once by hand on the first day and once a month after that. On the first run it researches the business from its own public surfaces before asking anything, reads the account structure once, writes the plan folder, seeds the board, creates the creative doctrine, builds the dashboard, and registers the recurring jobs. It asks about exactly two things, the monthly ceiling and the daily cap, because those are the two numbers that decide how much money can leave. On every monthly run it re reads the evidence, corrects what the ledgers prove wrong, and reconciles drift.
Monthly, on the last weekday. Folds a month of the creative ledger against the daily metrics ledger to score angle, format, hook, and offer framing rather than individual files, tests every line currently in the creative doctrine against what the month actually measured, and rewrites the doctrine where the evidence disagrees with the assumption. It records the fatigue curve it observed, retires what has not earned, and changes nothing where the month holds too few rows to tell one angle from another.
Runs once by hand on the first day and once a month after that. On the first run it researches the business from its own public presence, writes the charter with a source and a date beside every conclusion, discovers the AI Employees actually installed on this machine within a bounded search, writes the map every other routine in this kit reads, seeds the priorities and the watchlist once, builds the fleet dashboard, and registers this Employee's own seven jobs. On every monthly run it re-runs discovery and reconciles the map against what is really there. It never writes into another Employee's folder, never registers a job it does not own, sends only where you released the channel, spends only where you released it, and never touches a credential.
Monthly, on the last weekday, file work only, no browser at all. Folds the decision ledger, scores this Employee's own judgement against the metric files that measured it rather than against its own predictions, appends one outcome line per closed decision, retires a move proposed three times and never accepted, and rewrites the priorities where the evidence disagrees with them. It refuses to publish a calibration figure on too few closed decisions, changes nothing outside its own folder, proposes nothing to anybody, sends only where you released the channel, spends only where you released it, and never touches a credential.
Weekdays, file work only, no browser at all. Takes the single highest ranked open fault on the fleet and turns it into something the member can act on in a minute: the fault in one sentence, the first record that shows it, what changed around that date, three ranked causes each with its evidence line, what the failure has cost, and one paste ready correction line with the exact path it belongs in. It reads other Employees strictly read only, writes nothing outside its own folder, sends only where you released the channel, spends only where you released it, and never touches a credential.
Weekly, file work only, no browser at all. Reads everything the week produced, picks exactly three moves, and argues both sides of each one before the member reads a word of it, with every clause carrying a number or an observation from a file it names. It argues against its own top recommendation last, records a predicted effect and the metric that would show it so the monthly review can score it, and files each move onto the register. It sends only where you released the channel, spends only where you released it, never deploys, and never touches a credential.
Weekly, heavy browser lane, read only toward the world with no exception. Works the surfaces on its own watchlist through one contiguous browser phase, captures one dated observation per real change with a verbatim quote and the exact URL, refuses to report any change against a surface it has no baseline for, and writes one capped weekly page. It types nothing that changes anything, takes no action on LinkedIn ever, sends only where you released the channel, spends only where you released it, and never touches a credential.
Weekdays, file work only, no browser at all. Walks every AI Employee installed on this machine, reads each one's own run log strictly read only, classifies every routine as running, stopped loudly, stopped silently, or paused, ages the faults, folds the decision register, and writes the one short brief the member opens first. It never writes into another Employee's folder, sends only where you released the channel, spends only where you released it, and never touches a credential.
Weekly on a Wednesday, light browser lane. Turns the questions that keep coming back into the answer that only has to be written once. It reads the ticket ledger over the recurrence window, finds every theme that crossed its threshold, checks the member's own help center and the existing macros for an answer that already covers it, and writes a reusable macro and a help article draft for every theme that has neither. It publishes only where you released the channel an article, never edits a live help center page, sends only where you released the channel, spends only where you released it, and never touches a credential.
Weekdays, conditional browser lane. Reads the ticket ledger and the account, billing, and usage surfaces the member named, and decides which paying customers are about to leave. Every flag is a dossier carrying the trip wires that fired, the exact evidence behind each one, that account's whole history, and one suggested save with its cost and the screen it happens on. It reads billing screens and touches no control on them: a refund, a credit, a plan change, and a cancellation are named and never made.
Weekly, conditional browser lane, read only everywhere. Scores the fleet and the business for the week with a source beside every single number and no number anywhere without one, cites the Employee that owns a metric rather than recomputing it, refuses a rate below the floor, and writes one capped weekly page. It opens a browser only where the metric map names a live screen, never writes into another Employee's folder, sends only where you released the channel, spends only where you released it, and never touches a credential.
Weekdays, file work only, no browser at all. Reads every ledger, every queue file, and every run record since its cursor, turns the member's ticks into replied lines and closed cards, computes the first response and time to resolution clocks the Friday report is built from, folds the card inbox, re-renders the desk board, and writes the short morning brief the member opens first. It holds every outbound action unless you released the channel, and it never touches a credential.
Weekdays, conditional browser lane. Selects from the ticket ledger hardest first, meaning by severity and then by age within severity, and writes each customer a reply into a dated queue file the member reads, edits, and sends by hand. Where the right answer is a refund, a credit, a plan change, or a cancellation, it names the remedy, the exact amount, and the exact screen, and leaves the granting to the member. It sends only where you released the channel, never posts, never resolves a ticket, spends only where you released it, and never touches a credential.
Weekdays, heavy browser lane. Reads every support channel named in strategy/channels.md, meaning the mailbox, the helpdesk queue, the review and rating listings, the marketplace pages, and the forums where the member's products are discussed, and captures every new or changed item as one dated, sourced, severity graded line on the ticket ledger. It replies to nothing, reacts to nothing, marks nothing read or resolved, and takes no action of any kind on LinkedIn. It holds every outbound action unless you released the channel, and it never touches a credential.
Monthly on the last weekday, light browser lane. Re-tests every theme and every severity rule against a month of real ticket evidence and rewrites strategy/themes.md wherever the evidence disagrees with the assumption. It merges two themes the evidence shows are one, splits a theme whose tickets have two different answers, retires a theme with no volume, creates a theme for a cluster that kept landing unclassified, and rewrites a severity rule wherever the outcomes contradict it. Theme ids are never reused and never renumbered. It holds every outbound action unless you released the channel, and it never touches a credential.
Weekly on a Friday, conditional browser lane. Scores the week from the ledgers with a source beside every number and an estimate nowhere, reads the review and rating movement off the member's own listing screens, replays the browser flows the weekday routines depend on, and names the one product change that would have removed the most tickets this week. It sends only where you released the channel, never posts, never resolves anything, spends only where you released it, and never touches a credential.
Weekdays, file work only, no browser at all. Reads every ledger, queue file, and run record, turns yesterday's ticks into sent dates and closed cards, folds the card inbox, rewrites the launch board, and writes the short morning brief the member opens first. It holds every outbound action unless you released the channel, and it never touches a credential.
Runs once by hand on the first day and once a month after that, light browser lane. On the first run it researches the business from its own published surfaces before asking anything, writes the strategy folder, creates the ledgers, seeds the opening cards, builds the single page dashboard, reconciles the schedule table, and registers the eight recurring jobs. On every monthly run it re-reads a month of the kit's own evidence, applies what changed, adds a tab for a channel that gained volume, rebuilds, and reconciles every registered job against the table. It holds every outbound action unless you released the channel, and it never enters a credential.
Monthly, on the last weekday. Re-tests every segment in strategy/icp.md against the evidence in this kit's own ledgers, then rewrites the targeting itself where the evidence disagrees with the assumption the segment was written on. It records every change in strategy/CHANGELOG.md. It sends only where you released the channel, never re-queues anyone, never edits the contacted ledger, and never writes a number it did not read out of a named file.
Runs once by hand on the first day and once a month after that. On the first run it researches the business from its own public surfaces before asking anything, writes the strategy folder, seeds the launch board, builds the tailored command center, reconciles the schedule table, and registers the recurring jobs. On every monthly run it re-reads the evidence the kit produced, applies what changed, reconciles drift, and rebuilds. It holds every outbound action unless you released the channel, and it never enters a credential.
Weekdays, browser only when the card needs one. Takes the next ready cards off the launch board and does the work each one names: stages copy into the dashboard, fills a directory or press form and leaves it open in its tab, verifies a setup, packages a handoff, or researches its own next targets. It ticks its own card the moment it has verified the file that closes it. It submits only where you released the channel, sends only where you released the channel, spends only where you released it, and never touches a credential.
Weekdays. Turns the contactable people in the signal ledger into today's due first touches and follow ups, written into dated queue files the member reads, edits, and sends by hand, and optionally into the member's own mailbox as unsent drafts. It holds every outbound action unless you released the channel, and it never touches a credential.
Weekly, on a Friday, read only everywhere. Scores the week from the ledgers with a source beside every number, replays the browser flows the other routines depend on, names one thing to kill and one thing to scale, and files both as cards. It sends only where you released the channel, spends only where you released it, never touches a credential, and never writes a number it did not measure.
Claude Skills — questions
Answers built from the skills we actually parsed.