skills
AGENTS.md2 starsChanged 4 months agoMITView on GitHub
This repo publishes reusable agent skills — Claude Code, Codex, Cursor, OpenCode and friends. Skills live at skills/<category>/<skill>/ — categories are repo/, work/, docs/, meta/ — each self-contained: a SKILL.md (YAML frontmatter + body) plus optional assets. No runtime code ships except one file: scripts/resolve-config.sh, mirrored into every skill that reads the config. That is what test/ covers — the skill prose has no test/ suite, and none is wanted there. Skill quality is measured out of band instead, by…
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 33,315 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository25 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- compact-readmeSkill
- vhs-demoSkill
- write-contributingSkill
- write-docsSkill
- write-readmeSkill
- tituskirch-skills-configSkill
- validate-skillsSkill
- atomic-commitSkill
- merge-depsSkill
- prune-branchesSkill
- prune-commentsSkill
- pull-requestSkill
- releaseSkill
- update-depsSkill
- write-gitignoreSkill
- decideSkill
- handoffSkill
- issueSkill
- refine-issueSkill
- tldrSkill
- work-implement-queueSkill
- work-implementSkill
- work-review-queueSkill
- work-reviewSkill
The file
We split this AGENTS.md into the rules it gives an agent
2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 5
Every file named below is the source of truth for what it configures.
- Very specific: line 60
Create skills/<category>/<name>/SKILL.md, run pnpm skills:sync, commit as feat(<name>): add skill.
≈ “Common” is an estimate: at least 1 in 5 of the 42,283 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
CLAUDE.md
This repo publishes reusable agent skills — Claude Code, Codex, Cursor, OpenCode and friends. Skills live at skills/<category>/<skill>/ — categories are repo/, work/, docs/, meta/ — each self-contained: a SKILL.md (YAML frontmatter + body) plus optional assets. No runtime code ships except one file: scripts/resolve-config.sh, mirrored into every skill that reads the config. That is what test/ covers — the skill prose has no test/ suite, and none is wanted there. Skill quality is measured out of band instead, by the evals/ convention (skills/README.md): per-skill, with/without-baseline fixtures that are inert dev-only data a development tool runs, never test/, and stripped by packaging and skills:link.
Every file named below is the source of truth for what it configures. Read it rather than trusting a summary here — this file carries only what no file states outright.
Agent instruction files
CLAUDE.md and AGENTS.md are kept byte-identical. CLAUDE.md is what Claude Code reads; AGENTS.md is what vendor-neutral agent tools read — Codex, OpenCode, Cursor, Copilot, and whatever follows them. Two real files, deliberately not a symlink: not every tool resolves one.
After editing either file, copy it over the other — don't repeat the edit by hand:
cp CLAUDE.md AGENTS.md # or the reverse, whichever you just editedRetyping a change is exactly how the two drift; one reflowed line or reworded clause is enough. diff CLAUDE.md AGENTS.md must print nothing — pnpm test fails when it does. Fix it by letting one file win wholesale, never by merging them.
Commands
package.json has the full list. The ones with non-obvious behaviour:
| Command | Why it needs saying |
pnpm verify | The repo's gate — check + skills:check + typecheck + test, in CI's order, and the root verify key. CI runs the same commands as one step each, so one run reports every failure; test/ci-gate.test.ts fails if the two lists drift. |
pnpm check | Lint + format check only — a subset of the gate, not the gate. pnpm check:fix applies both fixers. |
pnpm skills:sync | Regenerates eleven artifacts from the skill folders. Run after touching any skill. |
pnpm skills:check | The CI guard for the above. Fails if any of them drifted. |
pnpm typecheck | tsc --noEmit. erasableSyntaxOnly is on, so an enum fails here, not at runtime. |
pnpm test | node --test over test/ — the resolver, the schema, skill self-containment, the CI-gate guard, the conformance-gate guard, the link destinations, and the CLAUDE.md/AGENTS.md mirror. |
pnpm skills:link | Symlinks every skill into both ~/.claude/skills/ (Claude Code) and ~/.agents/skills/ (Codex, Cursor, OpenCode, Gemini CLI) for live local testing — whole-folder, except a skill carrying a dev-artifact dir (evals/) links entry by entry to leave the fixture out, per destination. pnpm skills:unlink clears both; the destination list lives once, in scripts/skills-lib.sh (ADR-0016). |
pnpm skills:conformance | Validates every skill against the Agent Skills spec with skills-ref, pinned and run in Docker. Not in verify — the gate stays pnpm-only; this needs Docker. Its own workflow. |
pnpm containers:audit | Optional Docker + Trivy check, outside verify. Scans the current conformance image from the registry and the VHS renderer built from its template. Prints counts and JSON-report paths; exit 0 = no advisories, 1 = advisories, 2 = a check could not run. |
Eleven artifacts are generated — never hand-edit them: the root README.md skills table, each skills/<category>/README.md, .claude-plugin/plugin.json, skills.sh.json's groupings, the <skills-config> block mirrored into each skill that hosts it — with templates/resolve-config.sh copied into each skill that names it, which is every host plus the two queue skills that delegate the prose and still run the script (source: scripts/config-block.md and scripts/resolve-config.sh), the <skills-authority> / <skills-authority-reduced> author-authority block mirrored into each skill that reads third-party text (source: scripts/authority-block.md), the <skills-verify> / <skills-verify-isolated> check-command block mirrored into each skill that runs the repo's gate (source: scripts/verify-block.md), the <skills-worklock> single-flight-lock spec mirrored into the two work-loop unit skills — the two *-queue skills name their worker's REFERENCE for it instead, which is what lets them skip both this block and the config one (source: scripts/worklock-block.md, roster in test/isolation.test.ts), the <skills-plan> plan-presentation rule mirrored into each skill that puts a plan in front of a human (source: scripts/plan-block.md, roster in test/isolation.test.ts), the <skills-tldr> report-lead rule mirrored into each skill whose run ends in a report — a strict subset of the plan roster, since a skill leads with a result only where it closes by reporting one (source: scripts/tldr-block.md, roster in test/isolation.test.ts), and the <skills-forge> forge/host resolution rule mirrored into each skill that drives a forge — the same two *-queue skills name their worker's REFERENCE for this one too (source: scripts/forge-block.md, roster in test/isolation.test.ts). A new category also needs an entry in CATEGORIES in scripts/gen-skills.ts, or the sync fails loudly.
Non-obvious tooling
- pnpm is mandatory —
packageManagerStrict: trueplus thepackageManagerpin. npm/yarn are rejected outright. pnpm-workspace.yamlholds every pnpm setting. pnpm 10+ reads only auth/registry from.npmrc; anything else placed there is silently ignored. There is no.npmrc.- A release-age gate is active (
minimumReleaseAgeinpnpm-workspace.yaml). Packages published inside that window will not install, andpnpm tazereports them as "up to date" rather than as withheld — so "up to date" never proves nothing newer exists. - oxfmt formats markdown, JSON and YAML too, not just JS — see
lint-staged.config.js. Markdown is the point here; JS is incidental. oxlintandoxfmtare pinned exactly (no^), which drops them out of taze's default scope entirely — absent from its table, not reported as current.-lbrings them back in and composes with the mode (taze minor -w -l). Don't reach forlatestto move a pin: it spans majors and will happily propose a 0.x major or a prerelease.scripts/toolchain-versions.jsonowns the OpenCode pin used by the workflow and integration test. Runtime and CLI consumers read their owner rather than duplicating literals. Eval/test manifests are excluded from workspace discovery and taze; they are development data, never packages to update.- Husky runs lint-staged and commitlint on every commit. Don't
--no-verifyunless asked. - Conventional Commits are enforced. Scope by skill name when changing one skill:
feat(write-readme): …. - release-please cuts releases from
mainwithbump-minor-pre-major, so a pre-1.0 breaking change bumps the minor. Onlyfeat/fix/breaking reach the changelog — typing a user-visible change aschoreorrefactorsilently drops it from the release.
CI gotchas
.github/workflows/ is authoritative. Four behaviours that mislead if unknown:
- CI runs
on: pull_requestonly, so a commit pushed straight todevis never checked by CI. The AI work loop is configuredbranch:dev— it commits to the shared branch with no PR — which makespnpm verify, run locally before the push, the only automated gate between such a change and the release branch. That is why the rootverifykey is the full gate and not just lint plus format: nothing downstream would catch a broken test, a type error, or a drifted generated artifact until the rollup PR. - Every PR workflow skips drafts, and that is now load-bearing rather than incidental — a draft's check list is not a pass.
work-implementopens its PRs as drafts on purpose andwork-reviewmarks one ready only once it would accept, so CI is spent once per finished instead of once per push (The draft gate inwork-implement's REFERENCE). Each of the three therefore needs both halves: the draft gate andready_for_reviewin its triggertypes, or the un-draft fires nothing. All three gate on the job, so a draft reports oneskippingrow per gated job rather than an empty list — which is why anything reading those checks must discardskippingoutright instead of treating a non-failing list as green. - CodeQL runs on every PR to
main/dev, markdown-only ones included —codeql.ymlcarries nopaths:filter, so both matrix legs (actions,javascript-typescript) analyse every pull request; only the draft gate above holds it back. It is not the cheap no-op a path-scoped workflow would be, and the work loop's CI read depends on knowing which workflows a head actually triggers, so this is worth having right. - The skill conformance check is a separate workflow, scoped to
skills/**and its own script — so it does not run on a PR that touches neither, and, like everything else here, never on a push straight todev. It re-tiers the deliberatedisallowed-tools(ADR-0007) instead of failing on it;test/conformance-gate.test.tspins that list tovalidate-skills' prose. Running it locally needs Docker, which is whypnpm verifydoes not.
Adding a skill
Create skills/<category>/<name>/SKILL.md, run pnpm skills:sync, commit as feat(<name>): add skill. The frontmatter contract and layout live in skills/README.md; the full workflow in CONTRIBUTING.md.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

