Gemini Partner

Invoke the Gemini CLI as a collaborative AI partner for development tasks. Gemini runs in headless mode with full filesystem access to the current project.

When to Use Gemini

Use Gemini CLI when you need:

1. Large Context Analysis - Gemini has 1M token context (vs ~200k for Claude)

   • Analyzing entire codebases at once
   • Bulk documentation generation
   • Large refactoring analysis

2. Second Opinion - Alternative AI perspective

   • Architecture decisions
   • Code review
   • Debugging complex issues
   • Evaluating approaches

3. Web Research - Gemini has built-in web search

   • Researching libraries/frameworks
   • Finding implementation examples
   • Checking latest documentation

4. UI Generation - Design wireframes and components

   • HTML/CSS wireframes
   • React/Vue/Svelte components
   • Full page layouts

Headless Invocation Pattern

Basic syntax (Gemini 3 Pro):

gemini -y -o json -m gemini-3-pro-preview "<prompt>"

Flags:

• -m gemini-3-pro-preview: ALWAYS use Gemini 3 Pro (latest and most capable model)
• -y or --yolo: Auto-approve all file operations
• -o json: Return structured JSON output

Important:

• Always specify -m gemini-3-pro-preview to use the latest model
• Run from the project working directory so Gemini has filesystem context

JSON Output Format

Gemini returns structured JSON with:

{
  "response": "The AI's text response",
  "stats": {
    "models": { /* token usage */ },
    "tools": { /* tool call stats */ },
    "files": { "totalLinesAdded": 0, "totalLinesRemoved": 0 }
  }
}

Invocation Examples

1. Codebase Analysis (Large Context)

gemini -y -o json -m gemini-3-pro-preview "Analyze this entire codebase and provide:
1. Architecture overview
2. Key patterns used
3. Potential improvements
4. Security considerations

Focus on the src/ directory structure and how components interact."

2. Second Opinion on Approach

gemini -y -o json -m gemini-3-pro-preview "I'm implementing [feature]. My current approach is [description].

Review my approach and provide:
1. Strengths of this approach
2. Potential issues or concerns
3. Alternative approaches to consider
4. Your recommendation"

3. Web Research

gemini -y -o json -m gemini-3-pro-preview "Research the latest best practices for [topic] in 2025.

Include:
1. Current recommendations
2. Common pitfalls to avoid
3. Relevant libraries/tools
4. Code examples if applicable

Cite your sources."

4. Generate UI Component

gemini -y -o json -m gemini-3-pro-preview "Create a React TypeScript component for [description].

Requirements:
- Use Tailwind CSS for styling
- Include TypeScript types
- Make it accessible (ARIA labels, keyboard nav)
- Add JSDoc comments

Save to: src/components/[ComponentName].tsx"

5. Documentation Generation

gemini -y -o json -m gemini-3-pro-preview "Generate comprehensive documentation for this codebase.

Create:
1. README.md with project overview
2. API documentation for public functions
3. Architecture diagram description
4. Getting started guide

Save documentation to the docs/ directory."

6. Code Review

gemini -y -o json -m gemini-3-pro-preview "Review the recent changes in this codebase for:
1. Code quality issues
2. Potential bugs
3. Security vulnerabilities
4. Performance concerns
5. Suggestions for improvement

Be specific with file paths and line numbers."

Prompt Engineering Tips

1. Be specific about output location - Tell Gemini where to save files
2. Request structured responses - Ask for numbered lists, sections
3. Provide context - Mention relevant files or patterns to consider
4. Set constraints - Specify technologies, patterns, or limits
5. Ask for reasoning - Request explanations for recommendations

Gemini's Built-in Tools

Gemini CLI has these capabilities (auto-approved with -y):

Error Handling

• Non-zero exit code indicates failure
• Check stats.tools.totalFail for tool errors
• Stderr may contain MCP warnings (usually ignorable)

Collaboration Pattern

When working with Gemini as a partner:

1. Claude initiates - Formulate the request and invoke Gemini
2. Gemini executes - Performs analysis/generation with 1M context
3. Claude reviews - Evaluate Gemini's output and integrate
4. Iterate if needed - Refine with follow-up prompts

See Also

• prompts.md - Curated prompt templates
• examples.md - Real-world usage examples
• Docs/Gemini_CLI/ - Full Gemini CLI documentation