Gemini System — Research & Multimodal Specialist

Gemini CLI (gemini-3-pro-preview) is your research specialist with 1M token context.

상세규칙: .claude/rules/gemini-delegation.md

Context Management (CRITICAL)

서브에이전트 경유 권장한다. Gemini 출력은 커지기 쉽기 때문에.

| 상황 | 방법 | |------|------| |코드 기반 분석 | 하위 에이전트를 통해(권장) | | 라이브러리 조사 | 서브 에이전트를 통해 (권장) | | 멀티모달 | 서브에이전트 경유(권장) | | 짧은 질문 (1-2 문 답변) | 직접 호출 확인 |

Gemini vs Codex

| Task | Gemini | Codex | |------|--------|-------| |리포지토리 전체 이해|✓| | |라이브러리 조사|✓| | |멀티모달(PDF/동영상/음성)|✓| | |최신 문서 검색|✓| | |디자인 판단| |✓| |디버그 | |✓| |코드 구현 | |✓|

When to Consult (MUST)

| Situation | Trigger Examples | |-----------|------------------| | Research | "검색" "리서치" / "Research" "Investigate" | | Library docs | "라이브러리" "문서" / "Library" "Docs" | | Codebase analysis | "코드베이스 전체" / "Entire codebase" | | Multimodal | "PDF" "동영상" "음성" / "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.

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:

gemini -p "Brief question" 2>/dev/null

CLI Options Reference

# 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 Korean

Output Location

Save Gemini research results to:

.claude/docs/research/{topic}.md

This allows Claude and Codex to reference the research later.

Task Templates

Pre-Implementation Research

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

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

# 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