Review Skills

review-skills

Voice

Read ../persona.md at the start of this skill. The mad-scientist voice is canonical for all output.

Scope: local to this skill's execution only. Revert to session default voice once the final report is printed.

Language

Adapt all output to the user's language. Technical identifiers (file paths, skill names, CLI flags) stay in their original form.

What this skill does

Two-phase quality pass over all nuthouse plugin skills:

1. Structure — runs /audit to check conventions (frontmatter, sections, persona pointer, plugin.json). Reports critiques and warnings.
2. Quality — lets the user pick which skills to test with skill-creator:skill-creator (evals, benchmarks, description optimization). Processes them one by one.

The two phases are independent. If the user only wants structure: point them to /audit. If they only want quality: start at Step 3.

Step 1 — Preconditions

Verify we're in the nuthouse repo:

test -f .claude-plugin/marketplace.json && echo "ok" || echo "not nuthouse"

If not nuthouse: abort with "ce labo n'est pas le bon."

Step 2 — Structural audit

Tell the user: "phase 1 — j'inspecte la structure. on vérifie les formules."

Dispatch /audit by invoking the audit skill directly. Let it run and print its report. Do not suppress its output.

After the report, ask:

des critiques ou warnings ? veux-tu corriger ça avant de tester la qualité ?
  (y) oui — stop ici, corrige d'abord les critiques
  (n) non — continue vers la phase qualité

If (y): exit. Print "reviens quand le labo est propre. 🧪"

Step 3 — Discover plugin skills

Discover all plugin skills (exclude local .claude/skills/):

find . -path "*/skills/*/SKILL.md" \
  ! -path "./.claude/*" \
  ! -path "./_templates/*" \
  | sort

Group by plugin. Present the list:

skills disponibles pour review qualité :

  acid-prophet
    [1] audit-spec    acid-prophet/skills/audit-spec/SKILL.md
    [2] check-drift   acid-prophet/skills/check-drift/SKILL.md
    [3] write-spec    acid-prophet/skills/write-spec/SKILL.md

  git-gremlin
    [4] commit        git-gremlin/skills/commit/SKILL.md
    [5] pr            git-gremlin/skills/pr/SKILL.md

  ... (full list)

lesquels veux-tu tester ? (numéros séparés par virgule, ou "all")

Wait for the user's selection. Parse the response into a queue of (plugin, skill, path) tuples.

Step 4 — Quality review loop

For each skill in the queue, in order:

1. Print: "skill-creator sur <plugin>:<skill> — lancement."

2. Chain to skill-creator:skill-creator with this context:

   "Audit this existing skill at <absolute path to SKILL.md>. The skill is named <plugin>:<skill>. Skip the intent-capture interview — the skill draft already exists. Go straight to writing 2-3 test cases, running evals (with-skill vs baseline), and launching the eval viewer. After the user reviews outputs and you've iterated to a good state, run the description optimization loop."

3. Wait for skill-creator to finish its loop with this skill.

4. After it returns, print:

   `<plugin>:<skill>` — done.
   queue restante : <N-1> skill(s)
   
     (c) continuer — prochain skill : `<plugin>:<next-skill>`
     (s) stop      — assez pour aujourd'hui
   

5. If (c): continue with next skill in queue.

6. If (s): exit with final report.

Step 5 — Final report

review-skills — terminé

  Structural audit:  /audit (voir rapport ci-dessus)
  Skills reviewed:   <N> / <total selected>
    <plugin>:<skill> — done
    <plugin>:<skill> — done
    ...
  Skills pending:    <list if stopped early, else "none">

Exit with voice line: "les créatures ont été testées. bonne nuit au labo."

Never

• Run git commit, git push, or git rebase.
• Suppress the /audit output — let it print in full.
• Run skill-creator on multiple skills in parallel — they each need human review.
• Skip the structural audit phase if there are unresolved critiques — quality testing on structurally broken skills is a waste.
• Auto-select "all" without showing the user the list first.