---

name: gemini description: Shell out to Gemini CLI from Claude for any task. Use when asked to "pipe content to Gemini", "get Gemini perspective", "use gemini CLI", "run headless AI task", "parallel analysis with Gemini", or "get second opinion from Gemini". Useful for alternative perspectives, batch processing, or leveraging Gemini 2.5 Pro capabilities alongside Claude.

Gemini CLI

Table of Contents

• Overview
• When to Use This Skill
• Core Command Pattern
• Model Selection
• Command Patterns
• Prompt Engineering for Gemini
• Common Use Cases
• Combining Claude + Gemini
• Advanced Patterns
• Error Handling
• Quick Reference

Overview

Execute tasks using Google's gemini command-line tool to leverage Gemini 2.5 Pro alongside Claude. This skill enables headless (non-interactive) execution where content is piped to Gemini, processed with custom prompts, and results are saved directly to files.

When to Use This Skill

Invoke this skill when:

• Alternative perspective: Get Gemini's viewpoint on a problem Claude is solving
• Parallel analysis: Run Claude and Gemini analysis side-by-side for comparison
• Specific Gemini strengths: Leverage capabilities where Gemini excels
• Headless processing: Need non-interactive AI execution that saves to files
• Batch processing: Process multiple inputs through Gemini
• Second opinion: Validate Claude's analysis with Gemini's assessment

When NOT to use: Tasks Claude can handle directly without needing a second AI perspective.

Core Command Pattern

The fundamental pattern for using gemini CLI:

cat [input-files] | gemini "[PROMPT]" --model gemini-2.5-pro > [output-file]

Components:

1. cat [input-files] - Concatenate input content (optional if prompt-only)
2. | gemini "[PROMPT]" - Pipe to gemini with custom instructions
3. --model gemini-2.5-pro - Specify the model
4. > [output-file] - Redirect output to a file

Model Selection

Model	Use For	Speed
gemini-2.5-pro	Complex reasoning, deep analysis, multi-step tasks	Slower
gemini-2.5-flash	Quick processing, simple transformations, batch jobs	Faster

Specify with: --model gemini-2.5-pro or --model gemini-2.5-flash

Command Patterns

Single File Input

cat document.md | gemini "Summarize in 3 bullet points" --model gemini-2.5-pro > summary.md

Multiple File Input

cat file1.txt file2.txt | gemini "Find common themes" --model gemini-2.5-pro > themes.md

Prompt-Only (No Input)

gemini "Create a 10-item code review checklist" --model gemini-2.5-pro > checklist.md

Prompt Engineering for Gemini

Define Role and Task

gemini "You are an expert [ROLE]. [TASK]. Format as [FORMAT]." --model gemini-2.5-pro

Multi-Part Instructions

gemini "Task:
1. [First step]
2. [Second step]
3. [Third step]

Format: Markdown with headings" --model gemini-2.5-pro

Common Use Cases

Alternative Perspective

Get Gemini's viewpoint alongside Claude's:

cat analysis.md | gemini "Review this analysis. Identify gaps or alternative viewpoints." --model gemini-2.5-pro > gemini_perspective.md

Verify output: head -20 gemini_perspective.md to confirm format matches expectations.

Comparative Analysis

Compare two approaches:

cat approach_a.md approach_b.md | gemini "Compare these approaches. Create pros/cons table, recommend when to use each." --model gemini-2.5-pro > comparison.md

Verify output: Check comparison.md contains expected table structure.

Code Review

Review code with Gemini:

cat code_file.py | gemini "Review for: 1) Design patterns, 2) Potential issues, 3) Security concerns. Provide specific examples." --model gemini-2.5-pro > code_review.md

Sample Output (code_review.md):

## Code Review

### Design Patterns: 7/10
- Uses functional approach appropriately
- Consider adding type hints

### Potential Issues
1. No input validation on line 15
2. Missing error handling for empty lists

### Security Concerns
- None identified for this scope

Verify output: grep "##" code_review.md to confirm sections exist.

Research Synthesis

Synthesize multiple sources:

cat source1.md source2.md source3.md | gemini "Synthesize key insights. Create executive summary followed by detailed findings." --model gemini-2.5-pro > synthesis.md

Batch Processing

Process multiple files:

for file in *.md; do
  cat "$file" | gemini "Summarize this document" --model gemini-2.5-flash > "summaries/${file%.md}_summary.md"
done

Combining Claude + Gemini

Sequential Processing

Claude processes first, Gemini refines:

1. Claude creates initial analysis, saves to file
2. Gemini reviews and enhances, saves refinement
3. Compare both results

Parallel Analysis

Both analyze the same input:

1. Claude analyzes in conversation
2. Gemini analyzes same content, saves to file
3. Compare perspectives, synthesize insights

Division of Labor

• Claude: Interactive exploration, code generation
• Gemini: Batch processing, alternative perspectives

Advanced Patterns

Iterative Refinement

# First pass
cat input.md | gemini "Analyze..." --model gemini-2.5-pro > v1.md

# Second pass with feedback
cat v1.md | gemini "Review and improve this analysis..." --model gemini-2.5-pro > v2.md

Chaining with CLI Tools

cat document.md | gemini "Extract key points" --model gemini-2.5-pro | grep "##" > headings.txt

Environment Variables for Reusable Prompts

PROMPT="Analyze for security vulnerabilities. Format as Markdown."
cat code.py | gemini "$PROMPT" --model gemini-2.5-pro > security_review.md

Error Handling

If command fails:

1. Verify gemini CLI is installed: which gemini
2. Check file paths exist: ls -la input.md
3. Ensure prompt is properly quoted
4. Verify model flag: --model gemini-2.5-pro

If output is poor:

1. Make prompt more specific
2. Add role definition
3. Request structured format
4. Try gemini-2.5-pro instead of flash

Quote escaping:

gemini 'Say "Hello World"' --model gemini-2.5-pro

Quick Reference

Prompt Template:

"You are [ROLE].
Task: [WHAT TO DO]
Input: [WHAT YOU'RE ANALYZING]
Output: [DESIRED FORMAT]
Context: [CONSTRAINTS/AUDIENCE]"

Best Practices:

1. Always specify model with --model
2. Define role in prompt
3. Request structured output format
4. Use descriptive output filenames
5. Verify output with head or grep after generation