Gemini Cli

Gemini CLI

Leverage Gemini's 1M+ context window as your AI pair programmer within Claude Code workflows.

This skill teaches Claude Code how to use the official Google Gemini CLI (gemini command) to get second opinions, architectural advice, debugging help, and comprehensive code reviews. Based on production testing with the official CLI tool.

---

Table of Contents

1. Quick Start 2. When to Use Gemini Consultation 3. Installation 4. Using Gemini CLI 5. Model Selection: Flash vs Pro 6. Proactive Consultation Patterns 7. AI-to-AI Prompting Best Practices 8. Common Use Cases 9. Integration Examples 10. Troubleshooting & Known Issues

---

Quick Start

Prerequisites:

• Gemini CLI installed (npm install -g @google/gemini-cli)
• Authenticated with Google account (run gemini once to authenticate)

Basic Usage Patterns:

# Quick question (non-interactive with -p flag)
gemini -p "Should I use D1 or KV for session storage?"

# Code review with file context (using --all-files)
gemini --all-files -p "Review this auth implementation for security issues"

# Architecture advice using Pro model
gemini -m gemini-2.5-pro -p "Best way to handle WebSockets in Cloudflare Workers?"

# Pipe file content for review
cat src/auth.ts | gemini -p "Review this authentication code for security vulnerabilities"

# Interactive mode for follow-up questions
gemini -i "Help me debug this authentication error"

---

When to Use Gemini Consultation

ALWAYS Consult (Critical Scenarios)

Claude Code should automatically invoke Gemini in these situations:

1. Major Architectural Decisions

• Example: "Should I use D1 or KV for session storage?"
• Example: "Durable Objects vs Workflows for long-running tasks?"
• Why: Gemini provides complementary perspective, may prioritize different concerns
• Pattern: gemini -m gemini-2.5-pro -p "[architectural question]"

2. Security-Sensitive Code Changes

• Authentication systems
• Payment processing
• Data handling (PII, sensitive data)
• API key/secret management
• Why: Gemini 2.5 Pro excels at security audits
• Pattern: cat [security-file] | gemini -m gemini-2.5-pro -p "Security audit this code"

3. Large Refactors

• Affecting 5+ files
• Core architecture changes
• Database schema migrations
• Why: Fresh perspective prevents tunnel vision
• Pattern: gemini --all-files -m gemini-2.5-pro -p "Review this refactoring plan"

4. Stuck Debugging (2+ Failed Attempts)

• Error persists after 2 debugging attempts
• Stack trace unclear
• Intermittent bugs
• Why: Different reasoning approach may spot root cause
• Pattern: gemini -p "Help debug: [error message]" < error.log

5. Context Window Pressure (70%+ Full)

• Approaching token limit
• Need to offload analysis to Gemini
• Why: Gemini's 1M context can handle large code files
• Pattern: cat large-file.ts | gemini -p "Analyze this code structure"

OPTIONALLY Consult

6. Unfamiliar Technology

• Using library/framework for first time
• Experimenting with new patterns
• Why: Gemini may have more recent training data
• Pattern: gemini -p "Best practices for [new technology]"

7. Code Reviews

• Before committing major changes
• Pull request preparation
• Why: Catches edge cases and improvements
• Pattern: git diff | gemini -p "Review these changes"

---

Installation

1. Install Gemini CLI

npm install -g @google/gemini-cli

2. Authenticate

gemini

Follow the authentication prompts to link your Google account.

3. Verify Installation

gemini --version

Should show version 0.13.0 or higher.

4. Test Connection

gemini -p "What is 2+2?"

---

Using Gemini CLI

Core Command Patterns

Non-Interactive Mode (-p flag)

Best for Claude Code integration:

# Direct question
gemini -p "Your question here"

# With model selection
gemini -m gemini-2.5-flash -p "Quick debugging question"
gemini -m gemini-2.5-pro -p "Complex architectural decision"

# With file context via pipe
cat src/auth.ts | gemini -p "Review this code"

# With all files in current directory
gemini --all-files -p "Analyze project structure"

# With stdin input
echo "Error: Cannot connect to database" | gemini -p "Help debug this error"

Interactive Mode (-i flag)

For follow-up conversations:

# Start interactive session with initial prompt
gemini -i "Let's discuss the architecture"

# Interactive with model selection
gemini -m gemini-2.5-pro -i "Help me design a database schema"

YOLO Mode (-y flag)

Auto-accepts all actions (use with caution):

# Dangerous: Auto-executes suggested commands
gemini -y -p "Fix all linting errors"

⚠️ Warning: YOLO mode can execute commands without confirmation. Only use in trusted environments.

---

Model Selection: Flash vs Pro

gemini-2.5-flash (Default)

Characteristics:

• Fast response time: ~5-25 seconds
• Good quality for most tasks
• Lower cost
• Safe for general questions

Use For:

• Code reviews
• Debugging
• General questions
• Quick consultations
• When speed matters

Example:

gemini -m gemini-2.5-flash -p "Review this function for performance issues"

gemini-2.5-pro

Characteristics:

• Response time: ~15-30 seconds
• Excellent quality, thorough analysis
• Higher cost
• Best for critical decisions

Use For:

• Architecture decisions (critical)
• Security audits (thorough)
• Complex reasoning tasks
• Major refactoring plans
• When accuracy > speed

Example:

gemini -m gemini-2.5-pro -p "Security audit this authentication system"

gemini-3-pro-preview (Preview - November 2025)

Characteristics:

• Google's newest and most intelligent AI model (announced Nov 18, 2025)
• State-of-the-art reasoning and multimodal understanding
• Outperforms Gemini 2.5 Pro on every major AI benchmark
• Supports text, image, video, audio, and PDF inputs
• Response time: TBD (in testing)
• Higher cost (preview pricing)

Use For:

• Most complex reasoning tasks requiring deep analysis
• Critical architectural decisions on cutting-edge projects
• Advanced multimodal analysis (diagrams, videos, complex docs)
• Benchmark-critical applications
• When absolute best quality is required (and cost is secondary)

Example:

gemini -m gemini-3-pro-preview -p "Deep architectural analysis of this distributed system"
cat architecture-diagram.png | gemini -m gemini-3-pro-preview -p "Analyze this system design"

⚠️ Preview Status: Model is in preview. For production-critical decisions, consider using gemini-2.5-pro (stable) until Gemini 3 reaches general availability.

How to Choose

Quick question? → Flash
Security/architecture? → 3-Pro-Preview (cutting-edge) or 2.5-Pro (stable)
Debugging? → Flash (try Pro if stuck)
Code review? → Flash
Refactoring 5+ files? → 3-Pro-Preview or 2.5-Pro
Multimodal analysis? → 3-Pro-Preview

---

Proactive Consultation Patterns

Pattern 1: Architecture Decision

Trigger: User asks about technology choice Claude Action: Automatically consult Gemini for second opinion

# Claude runs:
gemini -m gemini-2.5-pro -p "Compare D1 vs KV for session storage in Cloudflare Workers. Consider: read/write patterns, cost, performance, complexity."

# Then synthesizes both perspectives

Pattern 2: Security Review

Trigger: Working on auth/payment/sensitive code Claude Action: Request Gemini security audit

# Claude runs:
cat src/auth/verify-token.ts | gemini -m gemini-2.5-pro -p "Security audit this authentication code. Check for: token validation, timing attacks, injection vulnerabilities, error handling."

Pattern 3: Debugging Assistance

Trigger: Error persists after 2 attempts Claude Action: Get Gemini's perspective

# Claude runs:
echo "[error message and stack trace]" | gemini -p "Help debug this error. What's the likely root cause?"

Pattern 4: Code Review

Trigger: Major changes ready to commit Claude Action: Request comprehensive review

# Claude runs:
git diff HEAD | gemini --all-files -p "Review these changes for: correctness, edge cases, performance, security, best practices."

---

AI-to-AI Prompting Best Practices

How Claude Should Format Prompts to Gemini

✅ GOOD: Context-Rich, Specific

gemini -m gemini-2.5-pro -p "I'm building a Cloudflare Worker with user authentication. Should I use D1 or KV for storing session data? Consider: 1) Session reads on every request, 2) TTL-based expiration, 3) Cost under 10M requests/month, 4) Deployment complexity."

❌ BAD: Vague, No Context

gemini -p "D1 or KV?"

Prompt Structure Template

[Context: What you're building]
[Question]
[Considerations: Key factors (numbered)]

Example: Architecture Decision

gemini -m gemini-2.5-pro -p "
Context: Building a real-time collaborative editing app on Cloudflare Workers.

Question: Should I use Durable Objects or Workflows for managing document state?

Considerations:
1. Need to handle WebSocket connections (100+ simultaneous users per document)
2. Document state must be consistent across all clients
3. Need to persist changes every 5 seconds
4. Budget: <\$100/month at 1000 documents
5. Simple deployment preferred
"

---

Common Use Cases

1. Technology Selection

# Compare two technologies
gemini -m gemini-2.5-pro -p "Compare Drizzle ORM vs raw SQL for Cloudflare D1. Consider: type safety, performance, query complexity, bundle size."

2. Security Audit

# Audit authentication code
cat src/middleware/auth.ts | gemini -m gemini-2.5-pro -p "
Security audit this authentication middleware. Check for:
1. Token validation vulnerabilities
2. Timing attack risks
3. Error handling leaks
4. CSRF protection
5. Rate limiting
"

3. Debugging Root Cause

# Analyze error logs
tail -100 error.log | gemini -p "
These errors started appearing after deploying auth changes. What's the likely root cause?

Context:
- Added JWT validation middleware
- Using @cloudflare/workers-jwt
- Errors only on /api/* routes
"

4. Code Review

# Review pull request changes
git diff main...feature-branch | gemini --all-files -p "
Review this pull request. Focus on:
1. Breaking changes
2. Edge cases not handled
3. Performance implications
4. Security concerns
"

5. Refactoring Plan

# Plan large refactor
gemini --all-files -m gemini-2.5-pro -p "
I want to refactor this Express app to Cloudflare Workers with Hono. Analyze the codebase and suggest:
1. Migration order (which routes first)
2. Potential blockers (middleware that won't work)
3. Testing strategy
4. Deployment plan
"

6. Performance Optimization

# Analyze performance
cat src/api/heavy-endpoint.ts | gemini -p "
This endpoint is slow (500ms+ response time). Identify performance bottlenecks and suggest optimizations.

Context:
- Cloudflare Worker
- Fetches data from 3 external APIs
- Processes 1000+ items
"

---

Integration Examples

Example 1: Claude Consulting Gemini Automatically

Scenario: User asks architectural question

User: "Should I use D1 or KV for storing user sessions?"

Claude (internal): This is an architectural decision. Consult Gemini for second opinion.

[Runs: gemini -m gemini-2.5-pro -p "Compare D1 vs KV for user session storage in Cloudflare Workers..."]

Claude (to user): "I've consulted Gemini for a second opinion. Here's what we both think:

My perspective: [Claude's analysis]
Gemini's perspective: [Gemini's analysis]

Key differences: [synthesis]
Recommendation: [combined recommendation]"

Example 2: Security Review Before Commit

User: "Ready to commit these auth changes"

Claude (internal): Security-sensitive code. Request Gemini audit.

[Runs: cat src/auth/*.ts | gemini -m gemini-2.5-pro -p "Security audit..."]

Claude (to user): "I've reviewed the code and consulted Gemini for security concerns:

Gemini identified: [security issues]
Additional checks I recommend: [Claude's additions]

Safe to commit after addressing: [list]"

Example 3: Debugging Assistance

User: "Still getting this error after trying your suggestions: [error]"

Claude (internal): Two failed attempts. Consult Gemini for fresh perspective.

[Runs: echo "[error details]" | gemini -p "Help debug..."]

Claude (to user): "Let me get a second opinion from Gemini...

Gemini suggests: [Gemini's diagnosis]
This makes sense because: [Claude's analysis]
Let's try: [combined solution]"

---

Troubleshooting & Known Issues

Issue 1: Not Authenticated

Error: Error: Not authenticated

Solution:

gemini
# Follow authentication prompts

Issue 2: Model Not Found

Error: Error: Model not found: gemini-2.5-flash-lite

Cause: Model deprecated or renamed

Solution:

# Use stable models
gemini -m gemini-2.5-flash -p "Your question"
gemini -m gemini-2.5-pro -p "Your question"

Issue 3: Rate Limit

Error: Error: Rate limit exceeded

Solution: Wait 1-5 minutes, then retry

Prevention: Space out requests

Issue 4: Large File Context

Error: File too large for context

Solution: Use --all-files carefully or pipe specific sections

# Instead of:
gemini --all-files -p "Review everything"

# Do:
cat src/specific-file.ts | gemini -p "Review this file"

Issue 5: Command Hangs

Cause: Interactive mode when expecting non-interactive

Solution: Always use -p flag for non-interactive commands

# ✅ Correct
gemini -p "Question"

# ❌ Wrong (will hang waiting for input)
gemini "Question"

---

Production Best Practices

1. Always Use -p for Automation

# ✅ Good for scripts
gemini -p "Question"

# ❌ Bad for scripts (interactive)
gemini

2. Select Model Based on Criticality

# Architecture/security → Pro
gemini -m gemini-2.5-pro -p "[critical question]"

# Debugging/review → Flash
gemini -m gemini-2.5-flash -p "[general question]"

3. Provide Context in Prompts

# ✅ Good
gemini -p "Context: Building Cloudflare Worker. Question: Best auth pattern? Considerations: 1) Stateless, 2) JWT, 3) <100ms overhead"

# ❌ Bad
gemini -p "Best auth?"

4. Pipe File Content for Reviews

# ✅ Good
cat src/auth.ts | gemini -p "Review for security"

# ❌ Inefficient
gemini -p "Review src/auth.ts" # Gemini has to read file separately

5. Handle Errors Gracefully

# Add error handling
if output=$(gemini -p "Question" 2>&1); then
  echo "Gemini says: $output"
else
  echo "Gemini consultation failed, proceeding with Claude's recommendation"
fi

6. Synthesize, Don't Just Forward

❌ BAD: Just paste Gemini's response

User: "Should I use D1?"
Claude: [runs gemini] "Gemini says: [paste]"

✅ GOOD: Synthesize both perspectives

User: "Should I use D1?"
Claude: [runs gemini]
"I've consulted Gemini for a second opinion:

My analysis: [Claude's perspective]
Gemini's analysis: [Gemini's perspective]
Key differences: [synthesis]
Recommendation: [unified answer]"

---

Version History

2.1.0 (2025-11-19):

• Added Gemini 3 Pro Preview model (gemini-3-pro-preview)
• Updated model selection guidance for Gemini 3
• Added multimodal analysis use cases
• Updated model comparison matrix

2.0.0 (2025-11-13):

• Complete rewrite for official Gemini CLI (removed gemini-coach wrapper)
• Direct CLI integration patterns
• Simplified to core use cases
• Updated command examples for gemini CLI v0.13.0+

1.0.0 (2025-11-08):

• Initial release with gemini-coach wrapper
• Production testing and experimentation
• 8+ documented errors prevented

---

Related Skills

• google-gemini-api - Gemini API integration via SDK
• google-gemini-embeddings - Gemini embeddings for RAG
• google-gemini-file-search - Managed RAG with Gemini

---

License

MIT - See LICENSE

---

Support

• Issues: https://github.com/jezweb/claude-skills/issues
• Email: jeremy@jezweb.net
• Official Gemini CLI: https://github.com/google-gemini/gemini-cli

gemini-cli

Status: ✅ Production-Ready | v2.1.0 | Last Updated: 2025-11-19

Use the official Google Gemini CLI directly from Claude Code for second opinions, architectural advice, debugging help, and comprehensive code reviews using Gemini's 1M+ context window. Now includes Gemini 3 Pro Preview (Nov 2025).

---

Auto-Trigger Keywords

This skill should activate when Claude Code encounters these scenarios:

Commands

• gemini (official CLI tool)
• gemini-cli (installation package)
• gemini-2.5-flash, gemini-2.5-pro, gemini-3-pro-preview (model names)

Use Cases

• "second opinion"
• "consult Gemini"
• "peer review"
• "architectural decision"
• "stuck debugging" (after 2+ attempts)
• "security audit"
• "compare implementations"
• "model comparison"
• "Flash vs Pro"
• "1M context window"
• "large codebase analysis"

Technologies

• Google Gemini
• Gemini CLI
• @google/gemini-cli
• AI pair programming
• AI-to-AI prompting

Scenarios (Proactive)

• Major architectural decisions (D1 vs KV, Durable Objects vs Workflows, etc.)
• Security-sensitive code (auth, payments, data handling)
• Large refactors (5+ files, core architecture changes)
• Context approaching 70%+ full
• Unfamiliar technology stack
• Debugging failures (2+ attempts on same bug)

---

When to Use

✅ Use This Skill When:

• [ ] User asks major architectural question ("Should I use D1 or KV for sessions?")
• [ ] Writing security-sensitive code (auth, payments, PII)
• [ ] Planning large refactor (5+ files or core architecture)
• [ ] Stuck debugging after 2+ failed attempts
• [ ] Context approaching 70%+ capacity, need full codebase analysis
• [ ] Unfamiliar with technology stack or library
• [ ] Need to compare two implementations (old vs new, alternative approaches)
• [ ] User explicitly requests Gemini consultation
• [ ] Want complementary perspective on critical decision
• [ ] Need 1M+ context window for comprehensive analysis

❌ Don't Use This Skill When:

• [ ] Simple questions answerable with Claude's knowledge
• [ ] Trivial code reviews (small, non-critical changes)
• [ ] Every single decision (reserve for critical scenarios)
• [ ] API rate limits exceeded
• [ ] No internet connection available

---

Quick Example

Scenario: User asks "Should I use D1 or KV for session storage?"

Claude's Workflow:

# 1. Detect architectural decision
# 2. Inform user: "Consulting Gemini for second opinion..."

# 3. Get Gemini's perspective (Pro model)
gemini -m gemini-2.5-pro -p "I'm building a Cloudflare Worker with 1000-5000 DAU. Should I use D1 or KV for session storage? Consider: 1) Session reads on every request, 2) TTL-based expiration after 30 days, 3) Remember me functionality, 4) Multi-device tracking, 5) Cost under $50/month."

# 4. Synthesize both perspectives
# 5. Present unified recommendation

User sees:

Based on consultation with Gemini:

**My Perspective**: Both viable, depends on priorities
- KV excels at read performance (sub-millisecond)
- D1 provides better querying for multi-device tracking

**Gemini's Perspective** (2.5-Pro): D1 recommended
- Strong consistency is critical for sessions
- Query flexibility for "remember me" across devices
- Cost effective at this scale

**Recommendation**: Start with D1

---

Installation

Via Marketplace (Recommended)

# Add marketplace
/plugin marketplace add https://github.com/jezweb/claude-skills

# Install skill
/plugin install gemini-cli@claude-skills

Manual Installation

# Clone repository
git clone https://github.com/jezweb/claude-skills.git
cd claude-skills

# Symlink skill
ln -s $(pwd)/skills/gemini-cli ~/.claude/skills/gemini-cli

Install Gemini CLI

# Install official Gemini CLI
npm install -g @google/gemini-cli

# Authenticate
gemini

# Test
gemini --version
gemini -p "What is 2+2?"

---

Features

Core Capabilities

• ✅ Direct CLI Integration: Use official gemini command (no wrapper needed)
• ✅ Model Selection: Flash (fast) vs Pro (thorough)
• ✅ 1M+ Context Window: Analyze large code files
• ✅ Non-Interactive Mode: Perfect for Claude Code automation (-p flag)
• ✅ File Context: Pipe files or use --all-files
• ✅ AI-to-AI Prompting: Optimized prompts for Gemini from Claude

Use Cases Covered

1. Architecture Decisions: Technology selection, pattern choices 2. Security Audits: Authentication, payments, sensitive data handling 3. Debugging Assistance: Root cause analysis after failed attempts 4. Code Reviews: Before commits, pull requests 5. Refactoring Plans: Large-scale changes (5+ files) 6. Performance Optimization: Bottleneck identification

Token Efficiency

• ~60-70% token savings vs manual consultation
• 6+ common errors prevented (see SKILL.md)
• Automated synthesis of both AI perspectives

---

Usage Patterns

Pattern 1: Quick Consultation

gemini -p "Should I use Drizzle ORM or raw SQL for Cloudflare D1?"

Pattern 2: Security Review

cat src/auth/middleware.ts | gemini -m gemini-2.5-pro -p "Security audit this auth middleware"

Pattern 3: Code Review

git diff HEAD | gemini --all-files -p "Review these changes for edge cases and security"

Pattern 4: Architecture Advice

gemini -m gemini-2.5-pro -p "Best way to handle WebSocket state in Cloudflare Durable Objects? Consider: 1) 100+ concurrent connections, 2) Persistence requirements, 3) Failover"

Pattern 5: Debugging

echo "[error message]" | gemini -p "Help debug this error. What's the likely root cause?"

---

Comparison: gemini-cli vs google-gemini-api

Use both: gemini-cli for consulting Gemini, google-gemini-api for building with Gemini.

---

Related Skills

• [google-gemini-api](../google-gemini-api/) - Gemini API integration via SDK
• [google-gemini-embeddings](../google-gemini-embeddings/) - Gemini embeddings for RAG
• [google-gemini-file-search](../google-gemini-file-search/) - Managed RAG with Gemini
• [openai-api](../openai-api/) - Alternative for second opinions

---

Known Issues

1. Model Deprecation: gemini-2.5-flash-lite removed (use flash or pro) 2. Rate Limits: Space out requests if hitting limits 3. Large Files: Use --all-files carefully (can exceed context)

See SKILL.md for complete troubleshooting guide.

---

Version History

2.0.0 (2025-11-13):

• ✨ Complete rewrite for official Gemini CLI (removed gemini-coach wrapper)
• ✨ Direct CLI integration patterns
• ✨ Simplified to core use cases
• ✨ Updated for gemini CLI v0.13.0+

1.0.0 (2025-11-08):

• Initial release with gemini-coach wrapper
• Production testing and experimentation

---

Contributing

Issues and improvements welcome:

• Issues: https://github.com/jezweb/claude-skills/issues
• Email: jeremy@jezweb.net

---

License

MIT - See LICENSE

---

Maintained by: Jeremy Dawes | Jezweb

Gemini Models Comparison Guide

Based on systematic testing (2025-11-08, updated 2025-11-19 for Gemini 3)

---

Available Models

gemini-3-pro-preview (Latest - Preview)

Status: 🆕 Preview release (November 18, 2025)

Characteristics:

• Google's newest and most intelligent AI model
• State-of-the-art reasoning and multimodal understanding
• Outperforms Gemini 2.5 Pro on every major AI benchmark
• Supports text, image, video, audio, and PDF inputs
• Response time: TBD (in testing)
• Higher cost (preview pricing)

Best For:

• Most complex reasoning tasks
• Critical architectural decisions requiring deep analysis
• Advanced multimodal understanding
• Cutting-edge capabilities and benchmarks
• When absolute best quality is required

Example:

gemini -m gemini-3-pro-preview -p "Complex architectural decision requiring deep analysis"
cat complex-diagram.png | gemini -m gemini-3-pro-preview -p "Analyze this system architecture diagram"

⚠️ Preview Status: Model is in preview. Use for evaluation and testing. Consider using gemini-2.5-pro for production-critical decisions until Gemini 3 reaches stable release.

---

gemini-2.5-flash (Default)

Characteristics:

• Fast response time: ~5-25 seconds (average ~20s)
• Good quality for most tasks: ⭐⭐⭐⭐
• Prioritizes: Performance, simplicity, speed
• Safe with directory scanning
• Lower cost

Best For:

• Code reviews
• Debugging (root cause analysis is strong)
• Directory/file scanning
• General questions
• When speed matters

Example:

cat src/auth.ts | gemini -p "Review this code"
echo "Error message here" | gemini -p "Help debug this error"

---

gemini-2.5-pro

Characteristics:

• Response time: ~15-30 seconds (average ~23s, often similar to Flash!)
• Excellent quality: ⭐⭐⭐⭐⭐
• Prioritizes: Correctness, consistency, thoroughness
• May get confused with directory scanning (tries to use tools)
• Higher cost

Best For:

• Architecture decisions (critical)
• Security audits (thorough)
• Complex reasoning tasks
• When accuracy > speed
• Major refactoring plans

Example:

gemini -m gemini-2.5-pro -p "Should I use D1 or KV for session storage? Explain trade-offs."
cat ./src/api/* | gemini -m gemini-2.5-pro -p "Perform a security audit on this code"

---

gemini-2.5-flash-lite

Status: ❌ Not accessible via Gemini CLI

Model exists in Gemini API but returns 404 error when accessed via CLI. Do not use.

---

When Models Disagree

Critical Finding: Flash and Pro can give opposite recommendations for the same question, and both can be valid.

Example (D1 vs KV for sessions):

• Flash: Recommends KV
• Prioritizes: Performance, edge caching, TTL
• "Usually acceptable" eventual consistency

• Pro: Recommends D1
• Prioritizes: Strong consistency, SQL queries
• "Critical" consistency for sessions

Why This Happens:

• Flash: Performance-focused
• Pro: Consistency-focused

How to Handle: 1. For critical/security decisions → Prefer Pro's perspective 2. For performance-sensitive apps → Consider Flash's perspective 3. For major architectural choices → Get both viewpoints:

   gemini -p "Question?"  # Flash (default)
   gemini -m gemini-2.5-pro -p "Same question"  # Pro

---

Model Selection Matrix

---

Performance Comparison

Based on testing with same question ("D1 vs KV for sessions"):

Key Finding: Pro isn't significantly slower than Flash on many queries!

---

Override Default Model

# Use Pro for single command
gemini -m gemini-2.5-pro -p "Review this code" < src/auth.ts

# Or set as environment variable for all commands in session
export GEMINI_MODEL=gemini-2.5-pro
gemini -p "Review this code" < src/auth.ts
gemini -p "Architecture question here"

---

Recommendations

Default Strategy: Use Flash for most tasks, Pro (2.5 or 3) for critical decisions

When to Consider Gemini 3 Pro Preview:

• Cutting-edge projects requiring latest capabilities
• Complex multimodal analysis (images, videos, PDFs)
• Most critical architectural decisions
• Benchmark-critical applications
• Evaluation of newest features
• When cost is not primary concern

When to Use Gemini 2.5 Pro (Stable):

• Production-critical decisions
• Security audits
• Major refactors
• Stable, proven performance required
• Cost-effective Pro-level analysis

When Flash is Better:

• Quick code reviews
• Debugging
• Directory/file scanning
• Non-critical questions
• Cost-sensitive workflows

When to Get Multiple Perspectives:

• Critical architectural decisions
• Technology choices that affect entire project
• Performance vs consistency trade-offs
• Compare preview vs stable recommendations

---

Last Updated: 2025-11-19 (Gemini 3 preview release) Source: Systematic testing documented in gemini-experiments.md + official Gemini 3 announcement

AI-to-AI Prompting Best Practices

The Gemini CLI skill uses AI-to-AI prompting format for effective collaboration (based on systematic testing)

---

Format Structure

[Claude Code consulting Gemini for peer review]

Task: [Specific task description]

Provide direct analysis with file:line references. I will synthesize your
findings with mine before presenting to the developer.

---

Why This Format Works

1. Prevents Role Confusion: Gemini knows it's advising Claude, not the human developer 2. Reduces Chattiness: More direct, less "helpful assistant" framing 3. Better Output: File:line references, concrete suggestions 4. Peer Review Dynamic: Two AI systems collaborating

---

Comparison

❌ Old Format (less effective):

You're an expert security researcher. Review this code for vulnerabilities...

• More verbose
• "Helpful assistant" tone
• Chattier responses

✅ New Format (recommended):

[Claude Code consulting Gemini for peer review]

Task: Security review - identify vulnerabilities...

Provide direct analysis. I will synthesize findings.

• Clear AI-to-AI context
• Direct task description
• Concise output

---

Custom Prompt Template

gemini "[Claude Code consulting Gemini for peer review]

Task: [Your specific task here - be concrete]

Provide direct analysis with [file:line | code examples | recommendations].
I will synthesize your findings before presenting to developer." \
--yolo -m gemini-2.5-flash

Key Elements: 1. [Claude Code consulting Gemini for peer review] - AI-to-AI context 2. Task: - Specific task 3. Provide direct analysis... - Output format 4. I will synthesize... - Workflow explanation

---

Source: Testing documented in gemini-experiments.md (Experiment 1)