Review skill

This skill reviews and fixes other skills — identifying issues and applying corrections with developer approval at each stage.

Stop after each stage and have changes reviewed with the user.

Note: The agent checks skills against conventions and best practices, then proposes fixes. The developer approves before changes are applied. When uncertain about intent, ask — don't assume.

See responsibilities.md for the full agent vs developer ownership matrix.

Each stage produces a checklist of findings (pass / issue / suggestion) and proposes fixes for any issues found. Stage 5 delivers an overall quality assessment and publish/revise/rethink recommendation.

0. Read and understand the skill (agent proposes, developer confirms)

   • Read the target SKILL.md at $ARGUMENTS (and any supporting files in the directory). If no argument was provided, ask the user which skill to review.
   • Summarize: what does this skill do, when is it invoked, and what workflow does it follow?
   • Confirm understanding before proceeding to review

1. Frontmatter review (agent reviews, then fixes with developer approval)

   Check that frontmatter is well-formed and follows conventions:

   • Is name kebab-case and does it match the directory name?
   • Is description present, specific enough to trigger correctly, and does it include when-to-use context?
   • Are invocation control fields appropriate for the skill's purpose?
     • Side-effect workflows should use disable-model-invocation: true
     • Background knowledge should use user-invocable: false
   • Is allowed-tools set if the skill should restrict tool access?
   • Is argument-hint present if the skill uses $ARGUMENTS?
   • Is the description written in third person? (descriptions are injected into the system prompt — third person reads naturally there)
   • Are context/agent set appropriately if the skill runs in isolation?
   • Are there unknown or misspelled frontmatter fields?

   Report findings as a checklist: pass / issue / suggestion. Then apply fixes for any issues, with developer approval.

2. Prompt structure review (agent reviews, then fixes with developer approval)

   Check the prompt body for structural quality:

   • Does it have a stop-after-each-stage instruction if it's a multi-stage workflow?
   • Is there a Stage 0 for understanding/confirmation before doing work?
   • Are agent vs developer responsibilities clear at each stage?
   • Does it use questions to guide analysis, not just imperatives?
   • Is the skill under 500 lines? Does it reference supporting files for detail rather than inlining everything?
   • Does the skill specify what artifacts or outputs it produces and in what format? (reports, checklists, files, modified documents)
   • Are file references at most one level deep from SKILL.md? (no file→file→file chains — Claude may partially read nested references)
   • Is the stage numbering consistent and logical?

   Report findings as a checklist: pass / issue / suggestion. Then apply fixes for any issues, with developer approval.

3. Effectiveness review (agent reviews, then fixes with developer approval)

   Check that the skill will work well in practice:

   • Are instructions unambiguous — will Claude interpret them correctly?
   • Is the scope right-sized? (Not trying to do too much in one skill)
   • Are $ARGUMENTS, $0, $1 used correctly if present?
   • Is dynamic context (exclamation-mark backtick syntax) used correctly if present?
   • If the skill uses $ARGUMENTS, does it handle missing or invalid arguments gracefully? (prompt the user, show usage, or fail with a clear message)
   • Does the skill address what happens when it encounters an error or unexpected state? (validation, recovery, or clear reporting — not every skill needs elaborate handling, but it shouldn't leave the developer guessing)
   • Do reference files over 100 lines include a table of contents so Claude can see their full scope?
   • Are supporting files referenced from SKILL.md?
   • Check for anti-patterns:
     • Overly broad or narrow description
     • No review pauses in a multi-stage workflow
     • Unreferenced supporting files in the directory
     • Task instructions in a skill with no context: fork and no clear action
     • Deeply nested file references (file→file→file chains that Claude may not follow)
     • Agent escape hatches: Soft language ("where appropriate", "if you have a clear fix", "as needed") that lets the agent skip required work. Completion gates must use hard prerequisites, not discretionary phrasing. Look for: conditional qualifiers on mandatory steps, missing definitions of done, and loops without explicit exit criteria.

   Report findings as a checklist: pass / issue / suggestion. Then apply fixes for any issues, with developer approval.

4. Alignment review (agent reviews, then fixes with developer approval)

   If the skill is intended for this repo's skill library, check alignment with existing skills:

   • Consistent formatting with other skills in the repo?
     • Blockquotes for philosophy notes
     • Bold for stage titles
     • Tables for comparisons
   • Includes "When to use this vs other skills" if it overlaps with existing skills?
   • Follows the composition model (flesh-out -> review-steps -> strong-edit -> agent-optimize)?
   • Has a responsibilities.md if it's a multi-stage workflow with mixed agent/developer ownership?

   If the skill is standalone (not for this repo), skip repo-specific alignment checks and note this.

   Report findings as a checklist: pass / issue / suggestion. Then apply fixes for any issues, with developer approval.

5. Summary and recommendations (agent leads)

   Provide a final summary:

   • Overall quality assessment (ready to use / needs minor fixes / needs rework)
   • Top 3 issues to address (if any)
   • Top 3 strengths (what the skill does well)
   • Recommendation: publish / revise / rethink
   • If the skill will be used across model tiers (Haiku, Sonnet, Opus), flag whether instructions are explicit enough for smaller models — what works for Opus may need more detail for Haiku

When to Use This vs Other Skills

Goal	Use
Review a document for polish	review-steps
Review a document for substantive critique	strong-edit
Review a SKILL.md for quality and conventions	review-skill
Create a new skill from scratch	Start with _template, then review-skill the result