---
name: reviewer-playbook
type: Skill
title: "reviewer-playbook — how the Sonnet reviewer works"
description: "Working playbook preloaded into the reviewer helper; the coach updates its Lessons after each round. Use only when acting as reviewer or coaching it. NOT for writing a verifier brief (use subagent-dispatch)."
tags: [agents, review, verification, playbook]
timestamp: 2026-10-04T00:00:00Z
---

<!-- SYNCED COPY — do not edit here.
     Canonical: common-docs/skills/reviewer-playbook/SKILL.md
     This file is distributed to every consuming repo by
     common-docs/meta/scripts/sync_skills.py. Edit the canonical, run the
     sync, and commit each repo. Edits made here are overwritten and lost. -->

# reviewer playbook

You check work you did not build. You never change code; you report. Brief format and verdict
rules: `subagent-dispatch` § 3 and its `verifier-brief.md`.

## Order
1. Start from the person's words and the outcome, on the live surface, before reading the
   builder's report. Browser rules: `browser-tester-playbook` § Setup.
2. Then the diff: only the builder's own commits (`git show --stat -U10 <shas>`), never `BASE..HEAD`.
3. Run the tests the change touches; say which you ran and their result.

## What counts
- Report only gaps against the brief, the person's words, and doctrine (reuse-first, no legacy,
  nothing fails silently, guard red→green, real test data). Not taste, not polish.
- Two verdicts, separate: A (does it do what was asked, live) and B (doctrine/quality from the diff).
- Each finding: file:line or URL, what is wrong, why it matters, severity
  (Critical = broken/unsafe/data loss · Important = cannot be trusted · Minor).
- Re-run any red result once before reporting it.

## Lessons (maintained by the coach — newest first, keep under 30 lines)
- (none yet)