review-skills

Review Skills

Audit every skill available in the current project context and produce a structured report covering conflicts, quality issues, and patterns worth spreading.

Process

1. Discover skills in scope

Locate the installed skills first — do not assume a location, check which ones exist. Installed skills live in different places depending on how they were installed, and scanning a single hardcoded path is how this audit silently reports zero skills:

for d in ~/.agents/skills ~/.claude/skills "$PWD/.claude/skills"; do
  [ -d "$d" ] && printf '%s: %s skills\n' "$d" "$(find "$d" -maxdepth 2 -name SKILL.md | wc -l)"
done

~/.agents/skills/ — installed via npx skills add, flat, one directory per skill. Usually the real set.
~/.claude/skills/ — global skills
<cwd>/.claude/skills/ — project-local skills installed into this project
Plugins: read ~/.claude/settings.json for enabledPlugins, then find their SKILL.md files under ~/.claude/plugins/cache/

If ~/.agents/.skill-lock.json exists, read it. It records each installed skill's source and skillPath in its origin repo, which lets you spot two problems a directory listing cannot: entries whose skillPath no longer exists upstream (the skill was renamed or moved, so npx skills update can no longer follow it), and installed copies that have drifted behind their source.

By default, audit what is installed — that is what actually loads. A source repo's skills/ directory is development-only. But when the user points you at a repo's skills/ directory, or asks about skills they are authoring, review the source instead and say which you reviewed. If both exist, compare them: a source repo that has renamed or restructured its skills while the installed copies still carry the old names is a real, silent failure — every cross-skill reference in the new generation dangles at runtime.

For each skill found, record:

Source: installed (~/.agents/skills), global, project-local, plugin, or source repo
Path: absolute path to SKILL.md
Name: from name frontmatter field
Description: from description frontmatter field
Body line count: lines in SKILL.md excluding frontmatter
Bundled resources: whether references/, scripts/, or assets/ exist alongside SKILL.md

Read every SKILL.md in full. Note referenced supporting files but only read them if needed to resolve an ambiguity.

2. Check for conflicts and overlaps

Compare all skill pairs. Flag pairs where:

Descriptions share trigger phrases or keywords that would make selection ambiguous for an overlapping user prompt
Names are similar enough to confuse (e.g. clean-code vs clean-comments)
Both claim the same domain without clear differentiation

For each flagged pair, describe which user prompts would be ambiguous and suggest how to differentiate the descriptions.

3. Evaluate each skill against the quality rubric

For each skill, check all 15 items below. Record pass/fail and a one-line note for any failing check. The purpose of these checks is to ensure skills trigger reliably and give the model clear, actionable guidance.

Size and structure 7. Body is under 500 lines 8. If body exceeds 200 lines, large reference content is split into separate files rather than inlined 9. Steps use numbered ### headers rather than walls of prose 10. No hardcoded absolute paths that break portability (~ and relative paths are fine)

Instruction quality 11. Uses imperative form ("Do X" not "You should do X" or "Claude will X") 12. Explains the why behind non-obvious instructions, not just what to do 13. Output templates use consistent formatting (XML-style tags or fenced code blocks, not a mix)

Resources 14. Every file referenced by name in SKILL.md actually exists on disk 15. If references/, scripts/ or assets/ directories exist, SKILL.md mentions them so the model knows to use them. This applies to model-facing resources only — files the skill reads or renders at runtime. evals/, skill-feedback-ledger.md and similar are development tooling the model never loads; requiring SKILL.md to mention them adds context noise and is not a finding.

Assign severity to each failing check:

Critical: blocks correct skill usage (missing description, non-existent referenced files)
Major: meaningfully degrades output quality (vague triggers, missing why, hardcoded paths)
Minor: polish issues (style inconsistency, slightly verbose)

4. Identify replicable good patterns

Look across all skills for patterns worth adopting more broadly. Signal good work, not just problems. Patterns to watch for:

Interview-before-action: skill gathers requirements from the user before producing output — makes output more accurate and reduces rework
GitHub issue as deliverable: skill creates a structured GH issue — creates a durable record and fits code-review workflows
Progressive disclosure: SKILL.md is lean, with detailed content in reference files — keeps context window clean
Vertical-slice philosophy: work broken into thin end-to-end slices rather than layers
Model override for simple tasks: model: haiku in frontmatter for lightweight, mechanical skills — saves tokens and latency
Bundled scripts for deterministic work: scripts/ directory for code the skill always runs — avoids the model reinventing it each time

For each pattern found, note which skills use it and which other skills could benefit from adopting it, with a brief explanation of why it would help.

5. Produce the audit report

Skills discovered: [N] total — [N] installed, [N] global, [N] project-local, [N] plugin Scanned: [which locations existed and were read; say if you audited source rather than installed]

Conflicts & Overlaps

[For each flagged pair:]

[skill-a] vs [skill-b]: [which user prompts would be ambiguous]
- Suggestion: [how to differentiate descriptions]

If none: No conflicts detected.

Quality Issues

Critical

[skill-name]: [rubric check #] — [one-line problem description]
- Fix: [specific actionable change]

Major

[same format]

Minor

[same format]

If none in a category: None.

Good Patterns Worth Spreading

[pattern name] — currently used in: [skill-a], [skill-b]
- Could benefit: [skill-x], [skill-y] — [brief explanation of how it would help]

Per-Skill Scorecards

[skill-name] ([source]) — [N] lines

Rubric: [N]/15 checks passed
Strengths: [1-2 bullet points]
Top fix: [single most impactful improvement, or "No issues found"]

[repeat for each skill]

Keep the report factual and concise. Cross-reference rather than repeat the same finding in multiple sections. If a skill passes all checks with no overlaps, say so briefly in its scorecard and move on.