agent-skills
AGENTS.md2 starsChanged 53 days agoMITView on GitHub
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 31,774 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 repository15 other files it gives its agents.
- be-msSkill
- commitSkill
- learnSkill
- okSkill
- quizzSkill
- refactorSkill
- reportSkill
- skill-creator-pjSkill
- skill-creatorSkill
- syncSkill
- test-driven-developmentSkill
- testSkill
- user-policy-pjSkill
- user-policySkill
- videoSkill
The file
We split this AGENTS.md into the rules it gives an agent
12 rules. About 1 is common, 2 are distinctive, and 9 are very specific to this project.
≈1
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
9
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 6
Every manifest skill is installed for Codex, Claude Code, and Cursor.
- Very specific: line 25
Add a skill: create skills/<name>/SKILL.md.
- Very specific: line 26
Remove a skill: delete its source and manifest entry.
- Very specific: line 27
Add a policy: add policies.<name> to manifest.yaml with source under shared-rules/ and either cursor (ready .mdc) or cursor_header …
- Very specific: line 31
Add a synced project: run import-project.sh, then commit and push the projects checkout.
- Very specific: line 32
Run ./tests/test-install.sh after changes to skills, manifest, policies, installer, uninstaller, sync, or tests.
- Distinctive: line 36
Write skills for agent execution: retain only text that changes an agent decision.
- Distinctive: line 38
Update repository context, conventions, or instruction files only on an explicit user request.
- Common: line 40
Never put secrets or personal data in tracked files.
- Very specific: line 42
Run uninstall.sh only on a copy of the repository, never on the checkout under work.
- Very specific: line 46
Keep README.md, AGENTS.md, and CONTRIBUTING.md under 100 lines each.
- Very specific: line 47
Use the commit skill for commits: stage only task hunks, split independent tasks, propose exact hunks and message, then wait for user …
≈ “Common” is an estimate: at least 1 in 5 of the 40,246 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
Repository Context
- Purpose: distribute portable skills and policies for Codex, Claude Code, and Cursor on Linux.
manifest.yamlis the install manifest. It is the source of which skills and policies are installed.skills/<name>/SKILL.mdis the canonical implementation of a skill.- Every manifest skill is installed for Codex, Claude Code, and Cursor. Read
manifest.yamlfor the current list; do not maintain a second list here. skill-creator-pjreusesskill-creatorwith project-local placement; generated project skills are not registered inmanifest.yaml; those under.claudesync only through the private projects checkout,.codex/skillsones do not sync.refactordelegates rule verification for every file type to a fresh-context, read-only subagent namedverify;skills/refactor/references/verify.mdis its verification procedure. No subagent configuration is installed.shared-rules/*.mdis the canonical shared policy;adapters/holds agent-specific representations.shared-rules/user.mdholds the user-level rules for every agent.lib.shholds the logic shared byinstall.sh,uninstall.sh, andsync.sh; add shared logic there, never duplicate it across scripts.install.shvalidates the manifest and installs selected entries by symlink or copy; policies are written as managed marker blocks inCLAUDE.md/AGENTS.mdand as Cursor rules.--pruneremoves symlinks of skills no longer in the manifest.--scope user|project|allselects user-level entries (default), project-level entries, or both; scopeprojectneeds no--agent. Scopeallskips a missing projects checkout; scopeprojectrequires it.uninstall.shreverses the install for the selected agents and scope and restores*.agent-skills-backup.*. With--agent alland a scope other thanprojectit also deletes this checkout.sync.shfast-forwards a cleanmaincheckout, fast-forwards the projects checkout when it is a Git repository, then runs the installer with--scope all --prune.- Project-level config (
.claude/contents of each project) lives in a separate private projects checkout,~/agent-skills-projects(override withAGS_PROJECTS_DIR):registry.yamlmapsprojects.<id>to the project'soriginURL, andprojects/<id>/.claude/holds the synced entries. This repository holds no project content. - The installer links (symlink by default, copy with
--mode copy;sync.shalways symlinks) each entry ofprojects/<id>/.claude/into<project>/.claude/for every matching project path. Runtime entries (is_project_runtimeinlib.sh, for examplesettings.local.json,jobs,worktrees) are never imported, linked, or pruned. - Project paths come from a per-machine, untracked
~/.config/agent-skills/local.yaml(override withAGS_LOCAL_CONFIG):roots(each root and its direct subdirectories) are scanned for repositories whose normalizedoriginmatches the registry,overridesmapid: pathfor projects without a remote or outside the roots, and every match expands throughgit worktree list. A project absent on a machine is skipped. import-project.sh --id <id> --path <project>copies a project's non-runtime.claudeentries into the projects checkout and registers itsorigininregistry.yaml; it never overwrites an existing entry and never commits or pushes.- Agent homes default to
~/.codex,~/.claude, and~/.cursor; skills install under each home'sskills/. Override homes withCODEX_HOME,CLAUDE_CONFIG_DIR, orCURSOR_CONFIG_DIR. - Manifest parsing in
lib.shsupports the repository's fixed indentation, not general YAML. Skill validation checks names, source paths, and file existence; it does not validateSKILL.mdfrontmatter or agent behavior. .github/workflows/test.ymlrunstests/test-install.shon Ubuntu for pushes and pull requests; tests use temporary homes, checkout copies, and local Git remotes.- This repository is public.
Change Routes
- Add a skill: create
skills/<name>/SKILL.md; add matchingskills.<name>.sourcetomanifest.yaml; add adapter or policy only when the target agent needs one. - Remove a skill: delete its source and manifest entry; update installation assertions.
install.sh --agent all --pruneremoves obsolete managed symlinks, including dangling ones; installed copies require separate removal. - Add a policy: add
policies.<name>tomanifest.yamlwithsourceundershared-rules/and eithercursor(ready.mdc) orcursor_header(frontmatter joined withsourceat install time); no script change. - Change a shared convention: update its canonical file in
shared-rules/, then update the required adapter. - Change installation, uninstall, sync, or manifest validation: update the scripts, keeping shared logic in
lib.sh, and extendtests/test-install.shfor the new behavior. - Change project-level sync: edit the
lib.shproject helpers (project_paths,project_targets,project_links,is_project_runtime),import-project.sh, and the--scopehandling in the three scripts; extendtest_projectsandtest_projects_sync. - Add a synced project: run
import-project.sh, then commit and push the projects checkout; other machines pick it up withsync.sh. - Run
./tests/test-install.shafter changes to skills, manifest, policies, installer, uninstaller, sync, or tests.
Conventions
- Write skills for agent execution: retain only text that changes an agent decision. Avoid explanatory prose, duplicate rules, and optional alternatives.
- Preserve one source of truth; do not copy policy text into unrelated files.
- Update repository context, conventions, or instruction files only on an explicit user request; do not infer that authorization from an implementation change.
- For a repository context update, read the actual implementation and update
AGENTS.md,CONTRIBUTING.md, andREADME.mdwhere their content is stale. A.docs/context/handoff alone does not satisfy that request. - Never put secrets or personal data in tracked files. Project-level config may hold internal content: commit it only to the private projects repository, never to this public repository.
- The first project install replaces real
.claudeentries and needs--force, which backs each one up;sync.shdoes not pass--force.uninstall.shrestores those backups only for the paths of the current scan. - Run
uninstall.shonly on a copy of the repository, never on the checkout under work; with--agent allit deletes its own directory. - The sync test builds its fixture with
git archive HEAD; commit script changes, or test a committed snapshot, before relying on it to exercise them.test_projects_syncoverlays the working tree onto its fixture clone instead. - Edits made to installed policy blocks and generated Cursor rules are overwritten on the next install; change the source in
shared-rules/instead. .docs/is local-only and must remain untracked.- Keep
README.md,AGENTS.md, andCONTRIBUTING.mdunder 100 lines each. - Use the
commitskill for commits: stage only task hunks, split independent tasks, propose exact hunks and message, then wait for user approval.
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.

