agents-skills
CLAUDE.md11 starsChanged 6 days agoMITView on GitHub
A public, shareable collection of AI toolchains — Skills and Agents — meant to be consumed by others across multiple AI coding tools (Claude Code, GitHub Copilot, Cursor). It is a content/config repo: there is no application code, no build step, and no test runner. "Correctness" here means each artifact is valid for its target tool and works when dropped into a consumer's environment. Because it is public and reusable, every artifact must be clean of personal and prior-project data.…
In this repository34 other files it gives its agents.
- AGENTS.mdAGENTS.md
- .github/copilot-instructions.mdCopilot instructions
- browser-diagnosisSkill
- cavemanSkill
- ci-flake-huntingSkill
- ci-speed-huntingSkill
- deploy-watchingSkill
- dispatching-subagentsSkill
- doc-coauthoringSkill
- docxSkill
- driving-prs-to-mergeSkill
- fast-forwarding-branchesSkill
- figmaSkill
- figma-specsSkill
- gh-issue-filingSkill
- gh-issue-labelsSkill
- gh-issue-lockingSkill
- grooming-issuesSkill
- humanizerSkill
- jira-issue-fieldsSkill
- jira-issue-filingSkill
- jira-issue-lockingSkill
- mcp-builderSkill
- orchestrating-slotsSkill
- pdfSkill
- pptxSkill
- pr-checksSkill
- pr-cleanupSkill
- pr-commentsSkill
- skill-creatorSkill
- spelunking-specsSkill
- technical-reconSkill
- token-bleed-huntingSkill
- xlsxSkill
The file
We split this CLAUDE.md into the rules it gives an agent
4 rules. About 0 are common, 2 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 22
When adding the first artifacts, create these directories.
- Very specific: line 31
When asked to add or rename a skill, treat the name as a public API: renaming breaks consumers who invoke it by name.
- Distinctive: line 67
Write artifact bodies in plain, tool-neutral language.
- Very specific: line 68
Keep frontmatter description action-oriented and specific — it is what each tool uses to decide when to trigger the skill/agent.
≈ “Common” is an estimate: at least 1 in 5 of the 27,641 CLAUDE.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 file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
What this repository is
A public, shareable collection of AI toolchains — Skills and Agents — meant to be consumed by others across multiple AI coding tools (Claude Code, GitHub Copilot, Cursor). It is a content/config repo: there is no application code, no build step, and no test runner. "Correctness" here means each artifact is valid for its target tool and works when dropped into a consumer's environment.
Because it is public and reusable, every artifact must be clean of personal and prior-project data. No employer names, internal URLs, ticket IDs, secrets, file paths from private repos, or learnings that only make sense inside one codebase. Keep learnings isolated — phrased generically so they apply anywhere.
Primary goal: extreme token savings at the highest practical accuracy. Every authored skill operates in caveman mode for terse working output, and model spend is tiered — the cheapest capable model per task, with mechanical work offloaded to a cheap-tier subagent and premium reserved for operator-gated hard reasoning. See MODEL-DEFAULTS.md.
Repository layout
Skills and Agents are the two primary artifact types. The canonical homes are the directories Claude Code auto-discovers:
.claude/skills/<skill-name>/SKILL.md— one directory per skill;SKILL.mdholds YAML frontmatter (name,description) + the skill body. Supporting files (references/, scripts) live alongside it..claude/agents/<agent-name>.md— one file per agent; frontmatter (name,description,tools, optionalmodel) + system prompt. Ships 18 tiered worker agents (scouton the cheap tier;builder,reviewer,implementer,pr-rescuer,diagnostician,verifier,spelunker,design-mapperon workhorse — these the commands dispatch — plus thepr-dispatcher,pr-comments,pr-checks,pr-cleanupsingle-lane PR watchers, thedeploy-watcherdeploy-health lane, and theci-speed-hunter,ci-flake-huntercontinuous CI lanes, thetoken-bleed-huntertoken-efficiency lane and thebranch-fffast-forward lane, all run via/loop). Canonical agents live here;make sync-agentsgenerates the Cursor (.cursor/agents/,model/readonlyschema), Copilot (.github/agents/*.agent.md), Kiro (.kiro/agents/,toolscategory allowlist), OpenCode (.opencode/agents/,mode/permission) and Codex (.codex/agents/*.toml,developer_instructions) copies — never hand-edit those..codex/config.tomlis hand-maintained. Keep agents thin and point at the relevant skill for the protocol..claude-plugin/marketplace.json+plugins/agents-skills/(Claude Code) and.cursor-plugin/marketplace.json+plugins/agents-skills-cursor/(Cursor) and.agents/plugins/marketplace.json+plugins/agents-skills-codex/(Codex/ChatGPT;plugin.jsonat the plugin root) — this repo doubles as its own plugin marketplace for all three ecosystems. Same skills in each (the Agent Skills format is identical); agents and commands come from that tool's generated mirror, since their frontmatter differs. Codex plugins carry no agents or commands, so the Codex payload is skills only. The plugin payload is generated from.claude/{skills,agents,commands}bymake build-plugin(real copies, not symlinks — Claude Code installs the plugin dir standalone, so a link escaping it would dangle);make check-pluginfails on drift. Never hand-editplugins/. Releases are cut by.github/workflows/release-plugin.yml: CalVerYYYY.M.D.N(patch counter so same-day releases never collide), tagv<version>, assets = one zip per plugin + one zip per skill (the latter for Claude Desktop and ChatGPT upload). The workflow stamps versions — never bump them by hand..claude/settings.json— repo-level Claude Code config. Currently pinsmodel: opusandincludeCoAuthoredBy: false(commits authored without the Claude co-author trailer).
When adding the first artifacts, create these directories. Top-level README.md is the human entry point.
GLOSSARY.md defines the vocabulary the artifacts assume: slot, dispatcher, watcher, hunting, ci-speed, flake, token bleed, plus lane, tick, primary checkout, and singleton. The watcher/hunter and speed/flake distinctions are load-bearing: several skills refuse work on the strength of them, so keep new artifacts consistent with these definitions.
Two kinds of skill
- Self-contained skills — the full instructions live here.
- Pointer skills — a thin
SKILL.mdthat points to an upstream project's skill (e.g. a plugin or another repo) so the latest version is always used. For these, the name defined here is the contract — keep it stable even when the upstream target moves. Document the upstream source and what the local name maps to inside the body; don't inline the upstream content (that defeats the point of pointing).
When asked to add or rename a skill, treat the name as a public API: renaming breaks consumers who invoke it by name.
Cross-tool consumability
The same artifact often needs to be reachable from six tools. Keep the content authored once (canonically under .claude/) and expose it to the others rather than forking copies:
- Claude Code — reads
CLAUDE.md,.claude/skills/,.claude/agents/natively. - GitHub Copilot — reads
.github/copilot-instructions.md(andAGENTS.md); agents in.github/agents/*.agent.md. - Cursor — reads
AGENTS.md(and.cursor/rules/*.mdcif you add project rules); agents in.cursor/agents/. - Kiro — reads agents from
.kiro/agents/<name>.mdand skills from.kiro/skills/(the Anthropic Agent Skills format, same as.claude/skills/). - OpenCode — reads agents from
.opencode/agents/<name>.mdand skills from.opencode/skills/(same Agent Skills format). - OpenAI Codex (CLI, IDE extension, ChatGPT desktop app) — reads
AGENTS.md; skills from.agents/skills/(same Agent Skills format); custom agents from.codex/agents/<name>.toml(TOML:name,description, optionalmodel,developer_instructions); custom prompts from~/.codex/prompts/only. The project.codex/layer (agents +.codex/config.toml) loads only once the project is trusted. ChatGPT web/mobile runs skills only, via the Codex plugin or per-skill upload — no custom agents or commands.
Because Kiro, OpenCode and Codex read the identical skill format, .kiro/skills, .opencode/skills and .agents/skills are symlinks to .claude/skills (one source of truth) — make sync-skills creates/repairs them. Agents, whose frontmatter differs per tool, are generated copies (see below). Prefer a shared AGENTS.md (Copilot + Cursor both honor it) over maintaining divergent copies. If a new artifact only makes sense for one tool, say so in its body.
Slash commands don't share a home or format — each tool needs its own copy, so the canonical .claude/commands/*.md are mirrored:
- Claude Code —
.claude/commands/<name>.md(YAML frontmatter:description,argument-hint,allowed-tools; body uses$ARGUMENTS). - Cursor —
.cursor/commands/<name>.md(plain markdown, no frontmatter; filename is the command; args = text typed after the command). - GitHub Copilot (VS Code) —
.github/prompts/<name>.prompt.md(frontmatterdescription/argument-hint/agent; body uses${input:args}). Copilot CLI does not support custom commands yet. - Codex —
.codex/prompts/<name>.md(frontmatterdescription/argument-hint;$ARGUMENTSnative; every other$token escaped as$$). Users copy them to~/.codex/prompts/and run/prompts:<name>. OpenAI has deprecated custom prompts in favour of skills.
The Cursor, Copilot and Codex copies are generated from the Claude source — after editing any .claude/commands/*.md, run make sync-commands (or python3 scripts/sync-commands.py) to regenerate them; make check-commands exits non-zero if they've drifted (wire it into CI / a pre-commit hook). Never hand-edit the mirrors. The script also prunes mirrors whose source command was renamed or deleted.
Model selection policy
Model spend is tiered and Anthropic-first for reasoning. Full task→tier matrix + per-platform model IDs: MODEL-DEFAULTS.md.
- Cheap / mechanical —
ls/grep/git status, file lookups, listing, fixed-command runs. Haiku 4.5, or each platform's cheapest native model (Copilot Raptor Mini, Cursor's free Auto). Always offloaded to a cheap-tier subagent that returns a distilled, caveman-compressed answer — never run in the main session. - Workhorse (default) — the orchestration loop, verification, dev work, filing/grooming/recon. Sonnet 4.6.
- Premium (operator-gated) — hard, cross-cutting reasoning only, never self-escalated. Opus 4.8.
Substitutes from other providers are acceptable only if they match on both axes: comparable context window (≈200K+, or 1M-class where the artifact needs it) and comparable reasoning tier — except the mechanical tier, which deliberately uses each platform's cheapest native model. State the substitution and why in the artifact when you make one.
Conventions
- Write artifact bodies in plain, tool-neutral language. Avoid assuming a specific local environment.
- Keep frontmatter
descriptionaction-oriented and specific — it is what each tool uses to decide when to trigger the skill/agent. - Validate by structure (valid frontmatter, required keys present, name unique) since there is no test suite.
- Token discipline for the orchestration family (
dispatch,rescue,gh-issue-*/jira-issue-*commands and their skills): operate in caveman mode to conserve tokens across the high-volume loop, but write the durable prose a human reads — issue/ticket bodies and PR descriptions — through humanizer so it doesn't read AI-generated. Under both, machine-precise content (labels/fields, JQL,gh/MCP commands, lock markers, theRECON OK/RECON-ERRORcontract,file:linerefs, code blocks, acceptance-criteria checklists) stays byte-exact.
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.

