Skill Library Health Report

Audits the set, not the skill.

Every skill review tool reads one file. None of them can see the problems that only exist in a collection — two skills claiming the same trigger phrase, four doing one job, a skill no wording will ever reach, siblings running different versions of the same framework. Each file, read alone, looks fine.

This reads all of them at once and reports what they do to each other.

MIT licensed. Free, complete, and not a sample of anything else.

---

Run it in ten seconds

python3 scripts/library_health.py

No arguments scans a bundled six-skill sample library and prints the full report, labelled as a demonstration. Nothing to configure, no keys, no network.

Run it on your own library

python3 scripts/library_health.py ~/.claude/skills/synced
python3 scripts/library_health.py ~/.claude/skills/synced --json

Point it at any folder containing skill folders. Nested trees are walked, so a plugins directory or a mixed layout works. Python 3, standard library only.

---

What it checks

| ID | Check | Question | |---|---|---| | H0 | Unreadable skill file | Did anything fail to load and get left out? | | H1 | Trigger collision | Do two skills claim the same phrase? | | H2 | Purpose overlap | Do several describe the same job? | | H3 | Unreachable skill | Is there anything for a request to match? | | H4 | Missing boundary | Does a skill with close neighbours say what it is not for? | | H5 | Description budget | Is the routing surface over the limit, or barely used? | | H6 | Stale or stub | Is the skill too thin to do what it claims? | | H7 | Version drift | Do siblings cite different versions of one framework? |

Full trigger rules, severity logic and the blind spot of every check are in references/health-checks.md.

What the score means

100 points, minus a deduction per problem — and the report prints the subtraction, not just the total:

  HOW THAT SCORE WAS REACHED
   100 starting points
      -8  6 colliding phrase(s)
     -20  35 closely related pair(s)
     -10  1 description over the 1024-character limit

A score you cannot reconstruct is a score you should ignore. Every weighting is a constant at the top of scripts/library_health.py — disagree with a number and change it deliberately.

A 90+ · B 75+ · C 60+ · D 40+ · E below 40.

What it refuses to do

• It will not grade a library of one. Four of the checks compare skills against each other and have nothing to compare, so no grade is issued and the report says why. An A awarded on an empty comparison is worse than no grade.
• A check that could not run is not reported as passed. Those come back N/A, listed separately from PASSED.
• It does not invent findings. A clean library scores 100 with "no deductions".
• It never hides a dropped input. A SKILL.md that fails to open is named in the report and counted in H0, because every other figure was computed without it.
• Similarity is a candidate for your judgement, never a verdict. Two skills that share vocabulary because they are one workflow split in two will show as an overlap. That is a finding to dismiss, not a defect to fix.

What it is not

Not a per-skill quality audit. Not a security scan. Not a judgement of whether any skill is well written, correct, or does what it claims. A library of excellent skills can score badly here, and a library of weak ones can score well.

Reads only

Executes nothing it scans. Writes nothing into your library. Makes no network call. The demonstration materialises its sample into a temporary folder and deletes it after the run — a tool that installed six fake skills to demo itself would be adding the problem it reports.

---

Two ways to use it

As a script — clone and run, as above.

As a skill — copy this folder into your skills directory. SKILL.md routes requests like "audit my skills library" or "why does the wrong skill keep firing" straight to it. A single-file build is in skill-library-health-STANDALONE.md for destinations that take one file only.

---

Verified

| Check | Result | |---|---| | monna-verify | CLEAN · 30/30 · 0 blockers · 0 fails · 0 AI-language hits | | Adversarial cold-run suite | 7 of 7 scenarios pass — cold run, score arithmetic reconstructed by hand, one-skill library, trigger collision, clean library, no-write/no-temp-residue, missing-library error | | Box rendering | every row exactly 80 columns across 1-, 3-, 6- and 125-skill libraries | | Real-world run | 125-skill library: grade E, 38/100, 6 colliding phrases, 35 close pairs |

---

From the same author

Built by MONNA Consulting. The paid portfolio covers what this one deliberately does not:

• Skill Security Auditor — eight checks on what a single skill could reach: injection surface, credential exposure, exfiltration paths, unbounded execution.
• Trade Post-Mortem Engine — behavioural fault classes in a closed trade log.
• Pre-Trade Risk Auditor — what a trade plan commits before it is sent.

This tool is not a trial of those. It answers a question they do not.

Contributing

Issues and pull requests welcome, particularly:

• additional check ideas with a stated blind spot
• false positives on real libraries, with the library shape that produced them
• weighting disagreements backed by a library where the current number reads wrong