Startproject

Phase 2: Requirements Gathering (Claude)

ユーザーに質問して要件を明確化。

Ask in Japanese:

1. 目的: 何を達成したいですか？
2. スコープ: 含めるもの・除外するものは？
3. 技術的要件: 特定のライブラリ、制約は？
4. 成功基準: 完了の判断基準は？

Draft implementation plan based on Gemini research + user answers.

Phase 3: Codex Design Review (Background)

Task tool でサブエージェントを起動し、Codex で計画レビュー。

Task tool parameters:
- subagent_type: "general-purpose"
- run_in_background: true
- prompt: |
    Review plan for: {feature}

    Draft plan: {plan from Phase 2}

    1. Call Codex CLI:
       codex exec --model gpt-5.4 --sandbox read-only --full-auto "
       Review this implementation plan:
       {plan}

       Analyze:
       1. Approach assessment
       2. Risk analysis
       3. Implementation order
       4. Improvements
       " 2>/dev/null

    2. Return CONCISE summary:
       - Top 3-5 recommendations
       - Key risks
       - Suggested order

Phase 4: Task Creation (Claude)

サブエージェントの要約を統合し、タスクリストを作成。

Use TodoWrite to create tasks:

{
    "content": "Implement {specific feature}",
    "activeForm": "Implementing {specific feature}",
    "status": "pending"
}

Phase 5: CLAUDE.md Update (IMPORTANT)

プロジェクト固有の情報を CLAUDE.md に追記する。

Add to CLAUDE.md:

---

## Current Project: {feature}

### Context
- Goal: {1-2 sentences}
- Key files: {list}
- Dependencies: {list}

### Decisions
- {Decision 1}: {rationale}
- {Decision 2}: {rationale}

### Notes
- {Important constraints or considerations}

This ensures context persists across sessions.

Phase 6: Multi-Session Review (Post-Implementation)

実装完了後、別セッションでレビューを実施。

Option A: New Claude Session

1. Start new Claude Code session
2. Run: git diff main...HEAD to see all changes
3. Ask Claude to review the implementation

Option B: Codex Review (via Subagent)

Task tool parameters:
- subagent_type: "general-purpose"
- prompt: |
    Review implementation for: {feature}

    1. Run: git diff main...HEAD
    2. Call Codex CLI:
       codex exec --model gpt-5.4 --sandbox read-only --full-auto "
       Review this implementation:
       {diff output}

       Check:
       1. Code quality and patterns
       2. Potential bugs
       3. Missing edge cases
       4. Security concerns
       " 2>/dev/null

    3. Return findings and recommendations

Why Multi-Session Review?

• Fresh perspective: New session has no bias from implementation
• Different context: Can focus purely on review, not implementation details
• Codex strength: Deep analysis without context pollution

User Confirmation

Present final plan to user (in Japanese):

## プロジェクト計画: {feature}

### 調査結果 (Gemini)
{Key findings - 3-5 bullet points}

### 設計方針 (Codex レビュー済み)
{Approach with refinements}

### タスクリスト ({N}個)
{Task list}

### リスクと注意点
{From Codex analysis}

### 次のステップ
1. この計画で進めてよろしいですか？
2. 実装完了後、別セッションでレビューを行います

---
この計画で進めてよろしいですか？

Output Files

Tips

• All Codex/Gemini work through subagents to preserve main context
• Update CLAUDE.md to persist context across sessions
• Use multi-session review for better quality assurance
• Ctrl+T: Toggle task list visibility