菜单
Review how an implementation diverged from its approved plan. Forks a clean-context agent over the plan and the diff to surface requirements drift and warranted follow-ups. Use at finish time or before opening a PR.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/bendrucker/claude --skill plan-review命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Working with GitLab merge requests via glab. Use when creating, updating, reviewing, or merging MRs, enabling auto-merge or merge trains, requesting or re-requesting reviewers, handling approvals, or working review threads, discussions, and draft notes. Load before running any `glab mr` or MR-mutating `glab api` command.
Create a pull request, merge request, or change request with proper formatting and content guidelines. Invoke when the user wants to create, open, or submit a PR, MR, or CR, including after committing changes.
Update a pull request or merge request body to reflect the current state of changes. Use when a PR/MR has evolved through additional commits and the body needs to reflect what will be merged.
基于 SOC 职业分类
正在显示 SKILL.md
| name | plan:review |
| description | Review how an implementation diverged from its approved plan. Forks a clean-context agent over the plan and the diff to surface requirements drift and warranted follow-ups. Use at finish time or before opening a PR. |
| allowed-tools | ["Agent"] |
Compare what shipped against the plan that was approved. Surface where the implementation drifted and what follow-ups that warrants.
The review runs in a forked agent with a clean context. An implementing session accumulates a justification for every departure as it happens, so drift goes unseen from inside it. A reviewer handed only the plan and the diff judges the delta cold. That outside view is the point: re-emphasizing the same prompt in the implementing context would not substitute for it.
Claude Code writes the approved plan to a file under ~/.claude/plans/ and injects that file's path into the session when plan mode exits. That file is the plan. Use the path already in this session's context. If none was injected, the session built against no approved plan and there is nothing to review: say so and stop.
Dispatch a background Agent: analyst if your available agent types list it, otherwise general-purpose. The review reads two documents and reports. It needs no tool an analyst lacks, and an analyst spawn arrives without the skill catalog a general-purpose one carries. Never fork: a fork inherits this session's context and defeats the outside view. The agent starts clean and has no idea where this skill lives, so hand it absolute paths: the plan file, and the rubric references/divergence.md resolved against this skill's base directory (that directory is in your context). A bare relative path resolves against the repo root in the agent's context and misses. Tell it to:
gh pr view --json baseRefName -q .baseRefName). With no PR yet, fall back to the repository default (gh repo view --json defaultBranchRef -q .defaultBranchRef.name). Don't assume main: a stacked branch targets its parent. Capture both committed and uncommitted work: git diff <base>...HEAD, plus git diff HEAD and git status --short for anything not yet committed. The review can run mid-implementation, so uncommitted changes are in scope.Run it on a cheaper model when one is set. The review reasons over two documents, so it does not need the top tier.
Relay the agent's report: the divergences and their classification, the drift, the ranked follow-ups (fix-before-merge against later), and the verdict. Add nothing of your own. The value is the outside view, not a second opinion filtered back through this context.