PR Review Skill
Updated 2026-10-01
What it is
A Claude Code plugin that runs a multi-persona PR review. Instead of one reviewer leaving a pile of comments, a sequence of specialized personas — an architect, a pragmatist, a security reviewer, an engineering manager, and optionally an implementer plus an independent verifier — work a pull request one after another, each reading what the others already said before adding its own.
The problem
A single reviewer, human or AI, tends to optimize for whatever it’s tuned toward — architectural purity, or shipping speed, or security — and only holds one of those concerns in mind at a time. The findings that actually matter in review are usually where those priorities disagree. /pr-review exists to surface that disagreement instead of averaging it away, and then to rule on it deliberately rather than leaving every thread open.
How it works
The pipeline is sequential and debate-shaped, not parallel: Architect (round 1) → 10x Engineer (round 1) → Architect (round 2) → 10x Engineer (round 2) → Security → Manager. Round 1 for the architect and the pragmatist can open new comment threads, capped by PR size — 3-5 new threads for a small PR, up to 12-20 for an XL one. Round 2 is response-only: they can concede or double down on what the other already said, but can’t raise new issues. Security and the manager have no comment cap.
Before the manager is allowed to render a verdict, a persona-dropout gate greps every comment already on the PR for the architect: / 10x: / security: prefix each persona is required to use, and aborts the whole pipeline if any one of them posted nothing — rather than letting the manager synthesize a confident verdict from a panel that’s actually missing a seat. Once the manager posts a Decision Summary (Must Fix / Should Fix / Consider / Defer) and the user picks a tier, a senior-engineer persona implements the selected fixes, and a separate verifier — with read-and-run access only, no commit or edit — independently re-checks each one: it greps the changed symbols to confirm they’re actually constructed and reached from the real composition root in production, not just referenced from a test. The fixes get pushed to the PR branch either way; the verifier’s finding only changes whether the final comment says the fixes are verified clean or lists what still needs a human look before merge.
Rule files for eight languages and five frameworks, each with a parallel security-focused file, are auto-detected from the diff and the repo and handed to the relevant personas — so a Go PR touching Orleans gets Go- and Orleans-specific guidance without anyone configuring it per repo.
Status
Stable, not currently under active development. Last commit was June 26, 2026.