One install gives your agent an audit of a skill you already have: it reads the real SKILL.md and its references, checks the description that carries the triggering, finds where two siblings claim the same request, catches stale names and counts, then fixes handoffs and command routing. Every finding quotes the line it came from, states a severity, proposes the smallest edit, and is validated by check_skill.py rather than by eye.
install
The install line from the repo README. It copies the whole folder, so references/ and scripts/ arrive with the skill.
npx skills add arm-pz/tune-skill --skill tune-skill
the audit
The audit works through references/audit-checklist.md. A finding needs evidence (the quoted line), a severity (breaks triggering, confusing, or cosmetic) and the smallest fix. The most severe is reported first.
The description is the only text present in every conversation, so it carries the triggering.
Two skills claiming the same request is the common fault. Overlaps only show when both descriptions are side by side.
"Pass along the findings" with no format is a broken handoff.
Stale names and counts, files that no longer exist, and prose doing a script's job.
Plain Python 3 plus PyYAML. It runs standalone on any skill folder and prints description length in characters, body in lines, so a run reproduces.
The exclusions are in the description, because that is the text that routes the request. Not for creating a skill from scratch: that is draft-skill. Not for turning a list of prompts into a skill: that is prompts-to-skill. When two skills claim one request, the fix is to give it one owner and narrow the other side to diagnosis plus a handoff offer, updating both descriptions in the same pass.
ground rules
READMEs, repo descriptions and listings drift from the SKILL.md the agent actually loads. Every claim about a skill comes from its SKILL.md or its references. If a file cannot be opened, the report names it and says what is being inferred instead.
Change what the audit found and nothing else, in the author's voice. Frontmatter keys it does not recognise stay. Never-invent, honesty, fairness and privacy rules stay at least as strong, and untouched text is copied exactly.
The report says which checks ran, which test cases were only written, and which files were never seen. Some agents ship an automated triggering loop; it does not run in a plain chat, so it is recorded as not run rather than implied.
Free, open, and every finding arrives with the quoted line it came from.
Install tune-skill Read the source