gemini-system

name: gemini-system description: | Gemini CLI leverages 1M context for three roles: codebase analysis, external research (Google Search grounding), and multimodal file reading. MUST use when PDF, video, audio, or image files need content extraction. Also use for large-scale codebase understanding and external research/survey. Auto-triggers: file extensions .pdf, .mp4, .mov, .mp3, .wav, .m4a. Planning/design → use Codex instead. metadata: short-description: Gemini CLI — 1M context analysis, research & multimodal

Gemini System — Analysis, Research & Multimodal

Gemini CLI leverages 1M context for large-scale analysis, research, and multimodal file reading.

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

Three Roles of Gemini

1. Codebase and Repository Understanding

Use Gemini's 1M context to analyze the entire project (when Claude's 200K is insufficient).

# Analyze project structure
gemini -p "Analyze this codebase: directory structure, key modules, patterns, dependencies, and architecture" 2>/dev/null

# Detailed analysis of specific files
gemini -p "Analyze this code: purpose, patterns, dependencies" < /path/to/file 2>/dev/null

2. External Research and Surveys

Use Gemini's Google Search grounding to research the latest information.

# Library research
gemini -p "Research: {library}. Latest version, features, constraints, best practices, pitfalls" 2>/dev/null

# Best practices research
gemini -p "Research best practices for {topic}. Latest recommendations, patterns, anti-patterns" 2>/dev/null

# Technology comparison
gemini -p "Compare {A} vs {B} for {use case}. Pros, cons, performance, community" 2>/dev/null

3. Multimodal File Reading

Extract content from PDF, video, audio, and image files.

# PDF
gemini -p "Extract: {what to extract}" < /path/to/file.pdf 2>/dev/null

# Video
gemini -p "Summarize: key concepts, timestamps" < /path/to/video.mp4 2>/dev/null

# Audio
gemini -p "Transcribe and summarize: decisions, action items" < /path/to/audio.mp3 2>/dev/null

# Image (diagrams, charts)
gemini -p "Analyze: components, relationships, data flow" < /path/to/diagram.png 2>/dev/null

Simple screenshot inspection can be done directly with Claude's Read tool.

Auto-Trigger

When multimodal files appear in a task, automatically pass them to Gemini without waiting for user instructions.

When NOT to Use Gemini

How to Use

Subagent Pattern (for large outputs)

Task tool parameters:
- subagent_type: "gemini-explore"
- prompt: |
    {task description}

    gemini -p "{prompt}" 2>/dev/null

    Save results to .claude/docs/research/{topic}.md
    Return CONCISE summary (5-7 bullet points).

Direct Call (for short extractions)

gemini -p "{what to extract/research}" 2>/dev/null

Language Protocol

1. Ask Gemini in English
2. Receive response in English
3. Report to user in the user's language