| name | meta-skill |
| description | Creates, improves, and reviews Claude Code Agent Skills. Use when the user wants to author a new skill, package a workflow or domain expertise into one, rightsize a skill that has grown too long, work out why a skill never fires, or health-check the whole skills directory to find which ones need work. Triggers on "create a skill", "make a skill for", "turn this into a skill", "improve this skill", "review my skill", "this skill never triggers", "why isn't my skill firing", "skills health check", "audit my skills", "which skills need work". |
meta-skill
A skill is a directory holding a SKILL.md and whatever it needs beside it. The frontmatter is
always in context, the body loads when the description matches, and the bundled files load only
when something reads them. Writing a skill is deciding what belongs at which of those levels.
Contracts
The rules you cannot derive from the files themselves. Everything else is your judgment.
- Frontmatter is required.
name is lowercase-hyphen and matches the directory the skill lives in.
- Personal skills live in
~/.claude/skills/, project skills in .claude/skills/ where git
shares them with the team.
- One skill is one capability. Two unrelated jobs are two skills.
SKILL.md is the only file loaded on every trigger. Every other path carries a condition for
opening it, and no instruction anywhere says to read one unconditionally.
- A bundled file that no route reads should not exist.
- Grade against
references/skill-review-rubric.md before calling a skill done.
Workflow
Select the single best-matching workflow and read its file for step-by-step instructions
before acting.
| Workflow | When to call it | File to read |
|---|
| Create Skill | The prompt asks for a new skill, or to package a workflow or expertise into one | workflows/create-skill.md |
| Improve Skill | The prompt asks to fix, shrink, review, or debug the triggering of a skill that already exists | workflows/improve-skill.md |
| Audit Skills | The prompt asks about the skills directory as a whole — a health check, which skills need work, what is broken | workflows/audit-skills.md |
Files
Paths below are relative to this skill's own directory — Claude Code announces it as
"Base directory for this skill" when the skill loads — not to the working directory.
| Path | What it is | Read it when |
|---|
templates/SKILL.md | Frontmatter and section skeleton for a new skill | Authoring a SKILL.md, on either route |
references/context-engineering-claude5.md | The Claude 5 rules for what to put in a skill and what to cut | Deciding what belongs in a skill, or running the subtraction test on a draft |
references/skill-review-rubric.md | Gradeable pass conditions, one per dimension | Grading a skill — last step of create, first step of improve |
scripts/scan.py | Mechanical pre-pass over every installed skill: what loads, what has a provable defect | Never — the Audit route executes it |
docs/claude_code_agent_skills.md | Upstream Claude Code skills guide | A Claude Code-specific format detail is genuinely in question |
docs/claude_code_agent_skills_overview.md | Upstream architecture doc, including cross-surface limits | A limit, or a surface other than Claude Code, is in question |
docs/blog_equipping_agents_with_skills.md | Anthropic's Skills launch post | The original progressive-disclosure argument is wanted in its own words |