| name | skill-audit |
| description | Deep audit of skill quality against current skill-creator best practices. Flags missing when_to_use, imperative density, missing evals/, and description budget overflow. |
| when_to_use | Use after writing or modifying a skill, before a release, or when reviewing skill quality. Triggers: "audit my skills", "check skill quality", "are my skills well-written", "/skill-audit", "skill best practice check", or after adding new SKILL.md files.
|
| user-invocable | true |
| argument-hint | [plugin-name] |
| allowed-tools | Bash Read Glob Grep |
/skill-audit — Deep Skill Quality Audit
Runs validate.js --audit to surface skill-quality issues that the standard /validate skips. Specifically:
- S09: Trigger coverage gaps — skills missing both
when_to_use and inline trigger phrases (combat undertriggering per skill-creator's "pushy descriptions" guidance)
- S10: Imperative density — skills with > 5
MUST/ALWAYS/NEVER occurrences (reframe stylistic imperatives as prose)
- S11: Missing evals/ — high-traffic skills (
feature-dev, bug-fix, cleanup, clarify, team-feature-dev, product-explore) without evals/evals.json (needed for skill-creator's run_loop.py description optimization)
- S12: Listing budget overflow — skills where
description + when_to_use exceeds the canonical 1,536-character cap
Plus everything /validate already checks (frontmatter validity, body length, model+effort assignment, antipatterns).
How to invoke
node "${CLAUDE_PLUGIN_ROOT}/scripts/validate.js" --root . --check skills --audit
node "${CLAUDE_PLUGIN_ROOT}/scripts/validate.js" --root . --check skills --audit --plugin nextc-core
node "${CLAUDE_PLUGIN_ROOT}/scripts/validate.js" --root . --check skills --audit --json
When to invoke automatically
The PreCommit hook (configured in ~/.claude/settings.json) blocks commits with WARN/FAIL findings. You can run /skill-audit proactively whenever you:
- Add a new skill
- Modify a SKILL.md description
- Increase a skill body's length
- Migrate to a new Claude Code spec version
Reading the report
Each finding has:
- Plugin name (e.g.
nextc-core)
- Skill name (e.g.
feature-dev)
- Check ID (S09–S12 for audit-mode checks)
- Level (PASS / WARN / FAIL)
- Message with concrete fix suggestion
Treat any FAIL as blocking. WARNs are advisory but in aggregate signal description budget waste, undertriggering risk, or skill-creator best-practice drift.
Reference
- Anthropic skill-creator:
~/.claude/plugins/marketplaces/claude-plugins-official/plugins/skill-creator/skills/skill-creator/SKILL.md
- Canonical frontmatter spec: https://code.claude.com/docs/en/skills
- Internal:
nextc-claude-toolbox/scripts/validate.js (audit-mode block at S09–S12)