Reviewing Skills

Reviewing Skills

Review a skill against Anthropic's skill-authoring best practices and report what could be improved. This is advisory: produce recommendations, never edit the skill unless the user explicitly asks. Nothing here blocks a commit.

The rubric you score against lives at references/skill-best-practices.md — a vendored, version-pinned copy of the upstream guidance. Read it before reviewing.

When to run this

• The user is creating or editing a SKILL.md, meta.yml, or a skill directory.
• The user asks "does this skill follow best practices?", "review my skill", or similar.
• Proactively, when you notice you're helping author a skill — offer a quick review.

Keep it lightweight and opt-in. Offer the review; don't force it, and don't gate the user's work on it.

Review workflow

1. Identify the target skill directory (the folder containing SKILL.md). If the user didn't name one, ask or infer from the file being edited.

2. Run the deterministic checks and capture the output:

   node scripts/review-skill.js <skill-dir>
   

   This reports the mechanical findings (body length, name/description rules, Windows-style paths, nested references, missing table-of-contents, rubric staleness). It is advisory and always exits 0. Fold its findings into your report rather than repeating the work by hand.

3. Read the rubric: references/skill-best-practices.md.

4. Assess the judgment calls the script can't — these are where you add value:

   • Is the description genuinely specific, or just superficially detailed? Would a different skill trigger instead?
   • Is the body concise, or does it explain things Claude already knows?
   • Is the degree of freedom right (prose vs. exact commands) for how fragile the task is?
   • Is progressive disclosure used well — overview inline, detail in linked files?
   • Are examples concrete (input/output pairs) where output style matters?
   • Is terminology consistent? Any time-sensitive content that will rot?
   • For skills with scripts: do they solve rather than punt, avoid magic constants, declare dependencies, and make execute-vs-read intent clear?
   • Is the naming clear (gerund form preferred, but noun-phrase and action-oriented are fine)? Only flag a name that is vague or genuinely confusing.

5. Emit the report using the template below.

Report template

Group findings by rubric area. Tag each as Pass, Suggestion, or Issue, and give a one-line concrete fix. End with a short, prioritized list. Keep it scannable.

# Skill review: <skill-name>

## Frontmatter
- [Issue] description is first-person — rewrite in third person: "Generates …".
- [Pass] name is valid and matches the directory.

## Conciseness & structure
- [Suggestion] SKILL.md body is 540 lines — move the API table into reference/api.md.

## Progressive disclosure
- [Pass] Detail files are linked one level deep from SKILL.md.

## Content & examples
- [Suggestion] The commit-message section would benefit from input/output example pairs.

## Scripts (if any)
- [Issue] scripts/run.py uses TIMEOUT = 47 with no rationale — document or derive it.

## Top recommendations
1. Rewrite the description in third person with explicit triggers.
2. Split the SKILL.md body below 500 lines.
3. Add two concrete examples to the commit-message section.

If there are no findings, say so plainly and note anything the skill does well.

Refresh mode

The vendored rubric carries a last_synced: date. When it's stale (the deterministic check flags it after 30 days), or when the user asks to update the best practices:

1. WebFetch the canonical URL recorded at the top of references/skill-best-practices.md.
2. Compare the fetched guidance against the vendored rubric and the checks in scripts/review-skill.js.
3. Propose edits to the rubric (and, if a mechanical rule changed, to review-skill.js) for the user to approve.
4. Update the last_synced: date in the rubric header.

Don't refresh silently — show the diff and let the user accept it.

Boundaries

• Advisory only. Do not edit the reviewed skill unless the user asks you to apply a fix.
• This skill complements, and never replaces, the blocking structural gate in scripts/validate-skills.js.