スキル一覧に戻る
mcsdodo

plan-review-skill

by mcsdodo

0🍴 1📅 2026年1月25日
GitHubで見るManusで実行

SKILL.md


name: plan-review-skill description: Use when about to implement a plan, when plan seems incomplete or has gaps, when unsure if tasks are feasible, or when validating spec before coding

Plan Review Skill

Two-phase review: analyze/document findings (separate agent), then apply approved changes (main context).

Quick Reference

PhaseWhereWhy
Phase 1: ReviewSeparate agentContext-intensive, autonomous
Phase 2: ApplyMain contextNeeds user interaction

Input: Plan path (e.g., _tasks/15-feature/02-plan.md) Output: _plan-review.md with findings + recommendation

Baseline Problem (Why Separate Agent)

Without separation: Main agent reads plan, creates review doc, manages iterations, tracks findings → consumes significant context before user interaction even begins.

With separation: Single Task tool call → agent returns summary only → main context preserved for Phase 2 user discussion.

Context savings: ~80% reduction in main conversation token usage for Phase 1.


Phase 1: Review → Separate Agent

Spawn single agent for entire Phase 1:

Task tool (general-purpose):
  description: "Plan review: {PLAN_NAME}"
  prompt: |
    Review {TARGET} and create {TARGET_DIR}/_plan-review.md.

    **Iterate (max 4)** until no NEW findings. Assess:
    - Completeness: requirements covered? edge cases?
    - Feasibility: achievable? hidden complexity? dependencies?
    - Clarity: implementer can follow? specific paths? verification steps?
    - YAGNI: unnecessary scope? duplication?

    **Checklist:** tasks have file paths, verification steps, correct order, no scope creep.

    Categorize: Critical/Important/Minor. Commit review. Return summary only.

After agent returns: Present summary to user, ask which findings to address/skip. STOP for user direction.


Phase 2: Apply → Main Context

After user approval only. Why main context? User interaction needed.

  1. Apply user-approved fixes to source plan
  2. Mark findings [x] in _plan-review.md
  3. Commit: git commit -m "plan: apply review feedback for {PLAN_NAME}"
  4. Update _plan-review.md with Resolution section (addressed/skipped)

Common Mistakes

MistakeFix
Running Phase 1 in main contextAlways spawn agent - preserves context
Applying fixes without user approvalSTOP after Phase 1, wait for direction
Skipping commit after reviewCommit _plan-review.md before presenting
Over-iterating (>4 rounds)Quality gate: stop when no NEW findings

Example

User: /plan-review _tasks/20-e2e-testing/02-plan.md
Claude: [Spawns Phase 1 agent]
Agent: "2 Critical, 3 Important, 1 Minor. Needs revisions."
Claude: Plan review complete. [summary] Full details: _plan-review.md
User: Address Critical and Important. Skip Minor.
Claude: [Phase 2 in main context - applies fixes]

スコア

総合スコア

60/100

リポジトリの品質指標に基づく評価

SKILL.md

SKILL.mdファイルが含まれている

+20
LICENSE

ライセンスが設定されている

+10
説明文

100文字以上の説明がある

0/10
人気

GitHub Stars 100以上

0/15
最近の活動

3ヶ月以内に更新がある

0/10
フォーク

10回以上フォークされている

0/5
Issue管理

オープンIssueが50未満

+5
言語

プログラミング言語が設定されている

+5
タグ

1つ以上のタグが設定されている

0/5

レビュー

💬

レビュー機能は近日公開予定です