Gemini System — Research & Multimodal Specialist
Gemini CLI (gemini-3-pro-preview) is your research specialist with 1M token context.
詳細ルール:
.claude/rules/gemini-delegation.md
Context Management (CRITICAL)
サブエージェント経由を推奨。 Gemini出力は大きくなりがちなため。
| 状況 | 方法 |
|---|---|
| コードベース分析 | サブエージェント経由（推奨） |
| ライブラリ調査 | サブエージェント経由（推奨） |
| マルチモーダル | サブエージェント経由（推奨） |
| 短い質問 (1-2文回答) | 直接呼び出しOK |
Gemini vs Codex
| Task | Gemini | Codex |
|---|---|---|
| リポジトリ全体理解 | ✓ | |
| ライブラリ調査 | ✓ | |
| マルチモーダル (PDF/動画/音声) | ✓ | |
| 最新ドキュメント検索 | ✓ | |
| 設計判断 | ✓ | |
| デバッグ | ✓ | |
| コード実装 | ✓ |
When to Consult (MUST)
| Situation | Trigger Examples |
|---|---|
| Research | 「調べて」「リサーチ」 / "Research" "Investigate" |
| Library docs | 「ライブラリ」「ドキュメント」 / "Library" "Docs" |
| Codebase analysis | 「コードベース全体」 / "Entire codebase" |
| Multimodal | 「PDF」「動画」「音声」 / "PDF" "Video" "Audio" |
When NOT to Consult
- •Design decisions (use Codex)
- •Debugging (use Codex)
- •Code implementation (use Codex)
- •Simple file operations (do directly)
How to Consult
Recommended: Subagent Pattern
Use Task tool with subagent_type='general-purpose' to preserve main context.
code
Task tool parameters:
- subagent_type: "general-purpose"
- run_in_background: true (optional, for parallel work)
- prompt: |
Research: {topic}
gemini -p "{research question}" 2>/dev/null
Save full output to: .claude/docs/research/{topic}.md
Return CONCISE summary (5-7 bullet points).
Direct Call (Short Questions Only)
For quick questions expecting brief answers:
bash
gemini -p "Brief question" 2>/dev/null
CLI Options Reference
bash
# Codebase analysis
gemini -p "{question}" --include-directories . 2>/dev/null
# Multimodal (PDF/video/audio)
gemini -p "{prompt}" < /path/to/file.pdf 2>/dev/null
# JSON output
gemini -p "{question}" --output-format json 2>/dev/null
Workflow (Subagent)
- •Spawn subagent with Gemini research prompt
- •Continue your work → Subagent runs in parallel
- •Receive summary → Subagent returns key findings
- •Full output saved →
.claude/docs/research/{topic}.md
Language Protocol
- •Ask Gemini in English
- •Receive response in English
- •Synthesize and apply findings
- •Report to user in Japanese
Output Location
Save Gemini research results to:
code
.claude/docs/research/{topic}.md
This allows Claude and Codex to reference the research later.
Task Templates
Pre-Implementation Research
bash
gemini -p "Research best practices for {feature} in Python 2025.
Include:
- Common patterns and anti-patterns
- Library recommendations (with comparison)
- Performance considerations
- Security concerns
- Code examples" 2>/dev/null
Repository Analysis
bash
gemini -p "Analyze this repository: 1. Architecture overview 2. Key modules and responsibilities 3. Data flow between components 4. Entry points and extension points 5. Existing patterns to follow" --include-directories . 2>/dev/null
Library Research
See: references/lib-research-task.md
Multimodal Analysis
bash
# Video gemini -p "Analyze video: main concepts, key points, timestamps" < tutorial.mp4 2>/dev/null # PDF gemini -p "Extract: API specs, examples, constraints" < api-docs.pdf 2>/dev/null # Audio gemini -p "Transcribe and summarize: decisions, action items" < meeting.mp3 2>/dev/null
Integration with Codex
| Workflow | Steps |
|---|---|
| New feature | Gemini research → Codex design review |
| Library choice | Gemini comparison → Codex decision |
| Bug investigation | Gemini codebase search → Codex debug |
Why Gemini?
- •1M token context: Entire repositories at once
- •Google Search: Latest information and docs
- •Multimodal: Native PDF/video/audio processing
- •Fast exploration: Quick overview before deep work
- •Shared context: Results saved for Claude/Codex