skill-reviewer

name	skill-reviewer
description	Reviews Agent Skill packages (SKILL.md, references, assets, tools) for structure, discoverability, safety, registry readiness, version discipline on publish, and conformance of body + folder to metadata.pattern (generator, reviewer, inversion, pipeline, tool-wrapper). Triggers: review skill, SKILL.md audit, pattern mismatch, wrong pattern, skill quality, skill.fish submit, agentskills format, bump version. Not for: application code (code-reviewer), pull requests (pr-reviewer).
metadata	{"pattern":"reviewer","severity-levels":"error,warning,info","version":1.2,"domain":"skills"}

Skill reviewer

Output contract: Deliver a review using the structure in references/review-output-template.md: Summary (verdict + score) · Findings grouped as error / warning / info · Top 3 recommendations · optional Registry note · Files reviewed checklist.

Rubric

Load and apply: references/skill-review-checklist.md.

Procedure

1. Scope — Skill directory (path to folder containing SKILL.md) or pasted SKILL.md + user-named paths. Out of scope: product/application code (code-reviewer), PR workflow (pr-reviewer).
2. Inventory — List SKILL.md, references/*, assets/*, tools/*, Folder_Structure.md if present. Note missing expected L3 folders when the body references them.
3. Passes (in order):
   • L1: name, description, triggers, metadata.pattern (canonical value?).
   • Pattern conformance: compare declared metadata.pattern to body + assets/ / references/ using checklist Pattern conformance table (generator → template, reviewer → rubric + severities, inversion → gates, pipeline → ordered steps, tool-wrapper → conventions + optional retriever). Flag error/warning when they disagree.
   • Layout: progressive disclosure, link depth, Folder_Structure.md.
   • Consistency: hybrid/secondary; guardrails placement; output contract vs pattern.
   • Safety: opaque shell, secrets, unvetted remote instructions.
   • Portability: monorepo subpath, skillfish, license if publish intent is clear; version bump (metadata.version incremented) if GitHub push / user-pack publish / submit is intended.
4. Findings — Each item: path or section · severity (error | warning | info) · why it hurts discovery, safety, or maintenance · concrete fix (not vague “improve X”).
5. Score — 1–10: 10 = checklist clean + clear output contract + publish-ready notes if applicable; deduct for each error and major warning.
6. Honesty — Do not invent file contents; if paths are missing, say what to open or create.

Pairing

• skill-creator — scaffold or rewrite skills after review.
• skillfish — submit, bundle, add, monorepo paths.
• Pack layout: LLM.md — Path rule (<pack-root>/<skill-id>/SKILL.md); enumerate skill folders under the pack root (no per-id list in LLM).

Guardrails

• Treat skills as unvetted for consumers unless source-reviewed; call out risky automation.
• Prefer actionable findings over boilerplate praise.