AgentSkillsCN

review-plan

当您希望在执行实施计划前对其进行关键性审查时使用。启动 Claude Code 子进程，对计划中的不一致之处、遗漏的边缘场景，以及架构层面的潜在隐患进行分析。

SKILL.md
--- frontmatter
name: review-plan
description: >-
  Use when you want a critical review of an implementation plan before executing
  it. Spawns a Claude Code subprocess to analyze the plan for inconsistencies,
  missing edge cases, and architectural concerns.

Review Plan

You MUST follow these steps exactly. Do NOT skip, reorder, or substitute any step.

Step 1: Locate the plan file

  • •If $ARGUMENTS is provided and non-empty, treat it as the plan file path. Resolve ~ to $HOME.
  • •If $ARGUMENTS is empty or not provided:
    1. •Use Glob to find all *.md files in ~/.claude/plans/.
    2. •The Glob results are sorted by modification time (most recent last). Pick the last file in the list.
    3. •If no files are found, tell the user: "No plan files found in ~/.claude/plans/. Provide a path: /review-plan /path/to/plan.md" and stop.

Step 2: Validate the plan file

Read the file to read the plan file. This is a validation gate — confirm the file exists and has content before spawning the subprocess. If the file does not exist or is empty, tell the user and stop. You do NOT need to analyze the content yourself; just confirm it's readable.

Step 3: Run the subprocess review

You MUST use Bash to run claude -p as a subprocess. Do NOT review the plan yourself — the entire point is to get an independent second opinion from a sandboxed subprocess.

Constructing the command

Use this exact pattern, substituting only the file path:

bash
claude -p "$(cat <<'PROMPT'
You are a senior staff engineer reviewing an implementation plan written by a
junior developer. Your job is to find problems, not to be encouraging.

Analyze the plan provided via stdin. Check for:

1. Logical inconsistencies or circular reasoning
2. Missing edge cases or failure modes
3. Overengineering or unnecessary complexity
4. Missing error handling at system boundaries
5. Unclear or ambiguous steps that could be misinterpreted
6. Missing verification steps or testing gaps
7. Architectural concerns (coupling, scalability, maintainability)
8. Security implications (injection, data exposure, auth gaps)
9. Dependency ordering problems (steps that assume prior work not yet done)

Output format — use EXACTLY this structure:

## Grade: X/5
A single number from 0 to 5, followed by a one-line justification.
- 0 — Critical errors that will cause failure if executed as-is
- 1 — Major problems that need significant rework
- 2 — Several important issues that must be addressed
- 3 — Solid plan with some gaps to fill
- 4 — Good plan, only minor improvements needed
- 5 — Totally safe to continue as written

## Overall Assessment
1-2 sentences. Be direct.

## Concerns
Number each concern. Tag severity as [Critical], [Important], or [Minor].
Explain WHY it is a problem and WHAT to do about it.

## What Works Well
Brief summary of strengths (2-3 bullet points max).
PROMPT
)" --output-format text --tools "" < PLAN_FILE_PATH

Replace PLAN_FILE_PATH with the actual absolute path to the plan file. Use input redirection (< file) rather than cat file | to avoid shell quoting issues with pipe.

Non-negotiable requirements

  • •--tools "" MUST be present. This sandboxes the subprocess so it cannot execute code or read files — analysis only.
  • •--output-format text MUST be present. This gives clean plaintext output.
  • •Do NOT add any other flags.
  • •Do NOT pipe with cat. Use < file input redirection.
  • •Set a Bash timeout of 120000 (2 minutes) since the subprocess needs time to analyze.
  • •If the subprocess returns an error or empty output, show the error to the user and suggest they try again.

Step 4: Present the results

Show the subprocess output to the user, prefixed with a brief note about which plan file was reviewed. Example:

code
Reviewed: ~/.claude/plans/gleaming-tumbling-allen.md

[subprocess output here]

Do NOT add your own commentary or re-analyze the plan. The subprocess output IS the deliverable.

Step 5: Capture session ID for follow-ups

The claude -p command prints a session ID to stderr (format: Session ID: ses_...). Capture it by appending 2>&1 or by extracting from the combined output.

After presenting the review, tell the user:

code
Session: <session-id>
To continue this review conversation, run: /review-plan --resume <session-id>

Resuming a previous session

If $ARGUMENTS starts with --resume, extract the session ID that follows and run:

bash
claude --resume SESSION_ID

This continues the existing review conversation so the user can ask follow-up questions, request clarification on specific concerns, or discuss fixes. Use --tools "" again to keep it sandboxed. Skip all other steps — go straight to this command and present the output.