Gemini CLI Skill

⚠️ Environment Notice

Environment	Command Format
Interactive terminal	gemini (enters interactive mode)
Claude Code / CI	gemini -p "prompt" (headless mode)
Scripting with JSON	gemini -p "prompt" --output-format json
Stdin input	echo "prompt" | gemini or cat file | gemini -p "analyze"

Non-TTY environments (Claude Code, CI pipelines) require -p flag or stdin input.

Quick Start

Headless Mode (Claude Code/CI)

# Basic review
gemini -p "Review this code for bugs"

# With JSON output for parsing
gemini -p "Analyze this code" --output-format json

# With specific model and directories
gemini -m gemini-3-flash-preview --include-directories ./src,./lib -p "Code analysis"

# Stdin input with prompt
cat src/auth.py | gemini -p "Review for security issues"

JSON Output Parsing

result=$(gemini -p "Query" --output-format json)
response=$(echo "$result" | jq -r '.response')

Reference Documentation

• Commands Reference - Slash commands, @ commands, shell mode
• Options Reference - Models, output formats, directories, JSON schema
• Examples - Code review, CI/CD integration, automation scripts

Available Models

Note: Model names change as Google releases new versions. Run gemini --help or check Gemini API Models for the current list.

Model	Description	Best For
gemini-3-flash-preview	Fast and efficient (DEFAULT)	Standard reviews, batch operations, general use
gemini-3.1-pro-preview	Latest flagship model	Complex architecture analysis, security audits

Output Formats

Format	Description	Use Case
(default)	Human-readable text	Terminal output
json	Structured with stats	Script parsing, automation
stream-json	JSONL events	Real-time monitoring

JSON Response Structure

Full schema: See Options Reference

Key fields: response (string), stats (object), error (optional object)

Key Options

Option	Alias	Description
--prompt	-p	Run in headless mode with prompt
--model	-m	Model selection
--output-format		Output format (json, stream-json)
--include-directories		Additional context directories (comma-separated)
--yolo	-y	Auto-approve all actions
--sandbox	-s	Sandbox mode (restrictive)
--approval-mode		Set approval mode (auto_edit)

Common Patterns

Full examples: See Examples for detailed patterns

Essential Patterns

# Code review with output
cat src/auth.py | gemini -p "Review for security issues" > review.txt

# JSON output with jq parsing
result=$(gemini -p "Query" --output-format json)
echo "$result" | jq -r '.response'

# Cross-directory analysis
gemini --include-directories ./backend,./frontend -p "Review API integration"

Timeout Configuration

Task Type	Recommended Timeout	Claude Code Tool
Quick checks	2 minutes	timeout: 120000
Standard review	5 minutes	timeout: 300000
Deep analysis	10 minutes	timeout: 600000

Recommendation: Use timeout: 600000 for complex analysis with gemini-3.1-pro-preview.

Error Handling

Detailed error handling patterns: See Examples

Error	Cause	Solution
No output	Missing -p flag	Use gemini -p "prompt"
Empty response	No stdin/prompt	Provide via -p or stdin
Exit code 1	General error	Check JSON .error field
Context too large	Too many files	Use specific paths
Permission denied	Sandbox restrictions	Use --yolo carefully

Best Practices

1. Use -p flag in Claude Code and CI environments
2. Use --output-format json for script parsing
3. Parse with jq for reliable extraction
4. Check .error in JSON response for error handling
5. Use --include-directories for multi-directory context
6. Match model to task: gemini-3-flash-preview for most tasks, gemini-3.1-pro-preview only for complex architecture/security
7. Set 10-minute timeout for deep analysis (timeout: 600000)