agentleFS
Sign inStart free
marcin-urbanskimarcin-urbanski/claude-skills/CLAUDE.md

claude-skills

CLAUDE.md0 starsChanged 7 days agoMITView on GitHub

This repo is the source of truth for my Claude skills (handoff, spec-first, design-check, tdd, systematic-debugging, verification-before-completion, ux-principles, pr, retro, playbook), the subagents spec-first delegates to (implementer, reviewer) and the handoff hooks (handoff-load.sh, handoff-nudge.sh). Files in this repo (skills, README, commits) are written in English. ~/Developer/claude-skills always stays on main. Do all changes in a separate worktree, so sessions never load a work-in-progress branch: Commit, push and open the PR from the worktree. After I merge the PR: It updates…

The file

We split this CLAUDE.md into the rules it gives an agent

7 rules. About 0 are common, 5 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    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
  1. Distinctive:

    Commit, push and open the PR from the worktree.

    line 19
  2. Very specific:

    Run bash tests/skill-frontmatter.test.sh: claude.ai rejects a description containing < or >, and Claude Code caps it at 1,024 characters.

    line 31
  3. Distinctive:

    Check the app's copy matches the repo:

    line 36
  4. Distinctive:

    Run each scenario at least 2 times with the new version, or 5 times when the change is only rewording, and compare with the baseline.

    line 53
  5. Very specific:

    When a skill misbehaves in real use but not in tests, read the real transcripts in ~/.claude/projects/<project>/*.jsonl (the folder name is …

    line 55
  6. Distinctive:

    If asked to publish all skills, list these exclusions with their reasons and ask before going further.

    line 62
  7. Distinctive:

    When adding, renaming or removing a skill or hook, update the tables in README.md.

    line 66

≈ “Common” is an estimate: at least 1 in 5 of the 26,980 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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
CLAUDE.md

claude-skills

This repo is the source of truth for my Claude skills (handoff, spec-first, design-check, tdd, systematic-debugging, verification-before-completion, ux-principles, pr, retro, playbook), the subagents spec-first delegates to (implementer, reviewer) and the handoff hooks (handoff-load.sh, handoff-nudge.sh). Files in this repo (skills, README, commits) are written in English.

How the repo is wired to Claude

  • ~/.claude/skills/<skill>, ~/.claude/agents/*.md and ~/.claude/hooks/*.sh are symlinks into ~/Developer/claude-skills. Every Claude Code session on this Mac loads whatever is checked out there, uncommitted edits included.
  • ux-principles has no local symlink. It reaches Claude through claude.ai sync. handoff is both symlinked and uploaded to claude.ai.
  • ~/.claude/skills/synced/ has not updated since 21.09 and has no effect on the app. Ignore it.

Branches and worktrees

~/Developer/claude-skills always stays on main. Do all changes in a separate worktree, so sessions never load a work-in-progress branch:

git worktree add ../claude-skills-<name> -b feature/<name> main

Commit, push and open the PR from the worktree. After I merge the PR:

cd ~/Developer/claude-skills && scripts/pr-done.sh <PR number>

It updates main, removes the worktree and deletes the branch. The change reaches sessions only after main is updated, because the symlinks point at ~/Developer/claude-skills. GitHub deletes the remote branch on merge. If pr-done refuses (uncommitted changes, or commits on the branch that are not in the PR), stop and tell me instead of cleaning up by hand.

After changing handoff or ux-principles

These two also live in claude.ai, so a change is not live there until re-uploaded.

  1. Run bash tests/skill-frontmatter.test.sh: claude.ai rejects a description containing < or >, and Claude Code caps it at 1,024 characters. Then build a ZIP whose top-level folder is named after the skill, from the checkout that has the change (the worktree, or ~/Developer/claude-skills after the pull):
    cd skills && zip -r ~/Downloads/<skill>.zip <skill> -x '*.DS_Store'
  2. I delete the old skill in claude.ai (Customize → Skills) and upload the ZIP.
  3. Check the app's copy matches the repo:
    ls -dt ~/Library/Application\ Support/Claude/local-agent-mode-sessions/skills-plugin/*/*/skills/<skill>
diff -r <newest path from above> skills/<skill>
    There are several skills-plugin/*/* folders, so the skill can show up more than once. Check the newest.

ux-principles/SKILL.md ends with a ## Sources section. Keep it: claude.ai gets the skill folder without the README, so the credits must live in the skill itself.

Testing skill changes

For skills where firing or a measurable behaviour matters (tdd, verification-before-completion), use claude plugin eval first: add or reuse a case in evals/ and follow evals/README.md. The no-plugin arm runs alongside, and the version on main is the baseline: run it before changing the skill, at least 5 runs per case. Before trusting a score, read a few kept traces (--keep-temp): a grader can be wrong too.

Skills that rest on judgment (spec-first, pr, retro) and changes no eval case covers are tested with subagents:

  1. Baseline first: run a subagent on each scenario the change targets with the current version (the one on main) and record what it does.
  2. Make the change in the worktree. The installed skill is still main's, so tell the test subagents to read the skill from the worktree path.
  3. Run each scenario at least 2 times with the new version, or 5 times when the change is only rewording, and compare with the baseline.

When a skill misbehaves in real use but not in tests, read the real transcripts in ~/.claude/projects/<project>/*.jsonl (the folder name is the project path with / replaced by -) to see what the model actually did.

What stays out of this repo

  • humanize-text: its sources' licences are unclear.
  • Emil Kowalski's animation skills (animate, find-animation-opportunities, improve-animations, mobile-native, review-animations in ~/.claude/skills): they are copies of emilkowalski/skills.

If asked to publish all skills, list these exclusions with their reasons and ask before going further.

Housekeeping

  • When adding, renaming or removing a skill or hook, update the tables in README.md.
  • Adapted skills keep their credit in README.md and THIRD_PARTY_NOTICES.md.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.

Check your setup, free