prompt-review

An engineering-grade review checklist for prompts, skills, and agent definitions. Runs the A-G checklist embedded in SKILL.md (each item tagged with severity and fix type) locally against the target file. Returns a REVIEW_PACKET: every issue carries severity, fix type (fix_type), the checklist item it hits, a line anchor, evidence, and acceptance criteria, plus a summary ledger. Issues with a canonical answer (direct) come with a concrete rewrite; issues that depend on the author's judgment (think) come with 2-3 questions rather than a fabricated answer. Whether to apply the fixes on disk, and whether to run more rounds, is left to the caller or the harness.

When to use it

• Right after writing a skill / agent / prompt file, when you want a line-by-line engineering critique
• Auditing an existing prompt library for common hard problems: hedging words like "should" / "try to", missing quantitative pass criteria, missing negative guardrails
• Reviewing a longer or multi-file skill / agent, one A-G dimension at a time
• Fully offline — the full checklist lives inside SKILL.md, self-contained, no network needed

How to trigger it

Say to Claude:

• "Review this prompt for me: agents/evaluator.md"
• "review skill ~/.claude/skills/foo/SKILL.md"
• "Audit this agent definition and check the prompt quality"

Or just paste a chunk of prompt text and ask for a review.

What you get back

• A local pass over the file against every A-G item; each hit becomes one issue
• A structured REVIEW_PACKET where every issue maps to an A-G item and carries a severity
• Each issue is split by fix type:
  • direct (canonical answer exists): a concrete rewrite / drop-in patch — e.g. replacing a hedge word, aligning a field name, adding a WRONG label
  • think (author judgment required): no fabricated answer — instead, 2-3 questions (what's the question / why only you can answer it / how would you rewrite once you decide), for cases like colloquial trigger phrasing, done criteria, or which of two conflicting hard rules should win
• A by_fix_type breakdown in the ledger so you can see at a glance how many are direct edits vs. how many need your own thinking
• This skill reviews only — it never writes to disk. Whether to apply fixes, and how many rounds to run, stays with the user or the harness.