メニュー
Use when the user asks to review a pull request, review a diff, review changes on a branch, or audit code written by a teammate or AI agent. Reviews the diff (not the summary) for unrelated changes, missing verification, behavior risks, and over-engineering.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Use when the user asks to fix a bug, investigate an error, debug a crash, explain a stack trace, or make a failing test pass. Enforces reproduce → isolate → smallest fix → verify, instead of guessing a cause and patching code.
Use when the user asks to add a new feature, add a new API/endpoint/page/component, implement a new capability, or extend existing functionality with new behavior. Forces a plan-before-code loop, scoping non-goals, and verifying the new behavior end-to-end.
Use when the user asks to understand, explain, trace, audit, or explore code WITHOUT modifying it. Examples — "how does X work", "where is Y called", "why does this branch exist", "trace the request flow". Stays read-only and does not produce a fix.
Use when a request is large, ambiguous, or spans multiple files/components and the user has NOT yet approved an implementation approach. Produces a verifiable plan via Plan Mode and a TodoWrite breakdown, then hands off to a more specific skill (bug-fix, refactor, feature-add) for execution.
Use when the user asks to refactor, clean up, simplify, modularize, deduplicate, rename, or improve the structure of existing code WITHOUT changing behavior. If the user wants new behavior, do not use this skill. Enforces small reversible steps and behavior preservation.
Use before releasing, deploying, publishing, cutting a tag, or merging into a production-bound branch. Classifies risk level and audits verification evidence, rollback plan, and operational readiness — not just whether the code compiles.
| name | acp-code-review |
| description | Use when the user asks to review a pull request, review a diff, review changes on a branch, or audit code written by a teammate or AI agent. Reviews the diff (not the summary) for unrelated changes, missing verification, behavior risks, and over-engineering. |
| version | 1.0.0 |
Use this skill when reviewing code written by humans or AI coding agents.
The goal is not to praise the implementation.
The goal is to find behavior changes, risks, missing verification, and unnecessary complexity.
Review the diff, not the summary.
An agent's explanation may be useful, but it is not evidence.
Evidence lives in:
Identify:
If the intent is unclear, ask before approving.
For each meaningful changed block, ask:
Why is this change necessary for the stated goal?
When the diff is large, prefer reading the actual git diff output directly rather than relying on the PR description. For cross-file impact (e.g. a renamed function), use the Explore subagent to find call sites.
Flag changes that are:
Look for:
A PR is weaker if it has no verification evidence.
Look for:
If verification is missing, ask for it.
Use one of these outcomes:
Summary:
- ...
Main risks:
- ...
Required changes:
- ...
Suggested improvements:
- ...
Verification gaps:
- ...
Be extra careful when the diff contains: