← ClaudeAtlas

pr-review-skill-generatorlisted

Generate a repo-specific PR-review skill by mining recently merged PRs — review comments, enforced corrections, accepted patterns, domain vocabulary, and post-merge fixes — into a distilled reviewer skill with provenance and a measured backtest score. Use when asked to generate, train, build, regenerate, or update a PR review skill from a repository's history.
paanSinghCoder/pr-review-skill-generator · ★ 0 · Code & Development · score 70
Install: claude install-skill paanSinghCoder/pr-review-skill-generator
# PR-review-skill generator You will mine a repository's merged PRs for the knowledge a senior reviewer carries in their head — what the team actually enforces, what it has deliberately accepted, its domain vocabulary and invariants — and distill it into a `pr-review` skill for that repo. The pipeline: ``` fetch (script) → select holdouts → mine (parallel subagents) → distill → write skill → backtest → report ``` Run phases in order. Each phase names the reference doc to read **when you reach it** — don't read them all upfront. ## Phase 0 — preflight 1. Check `gh auth status` succeeds and you're inside the target repo (or the user named one). This tool requires GitHub; if the repo is on GitLab or Bitbucket, say v1 doesn't support it and stop. 2. Detect an existing generated skill at `.claude/skills/pr-review/SKILL.md`. If present, read its `trained-range` metadata and switch to **incremental mode** (see below). 3. Ask the user (one round of questions, then proceed without further asking): - How many merged PRs to mine? Default **100** (cost scales linearly). - **Full run or preview first?** A preview mines ~15 PRs, shows the top candidate rules and the would-be decisions.md entries, then stops and asks before the full run (no skill written, no backtest). Recommend the preview to first-time users. - Output location? Default `.claude/skills/pr-review/` in the target repo. - Monorepo scoping — restrict to a subdirectory, or mine everyt