Gemini Review Skill

Overview

This skill enables a dual-AI review workflow where Claude prepares context and Gemini provides independent validation. The output always includes both Gemini's raw feedback and Claude's synthesized action plan.

Model: gemini-3-pro-preview (fixed) Cost: Free tier (60 requests/min, 1000 requests/day)

When to Use

• Validating implementation plans before coding
• Reviewing completed code for bugs, security, and best practices
• Discussing architecture decisions with a second AI perspective
• Getting specialized feedback for FastAPI or Next.js projects

Workflow Types

Type	Purpose	When to Use
plan	Validate implementation plan	Before starting development
code	Review completed code	After implementing features
architecture	Discuss system design	During design phase or refactoring

Execution Process

Step 1: Gather Context from User

Before executing, collect:

1. Review type: plan, code, or architecture
2. Tech stack (optional): fastapi, nextjs, or general
3. Content to review: The actual plan, code, or architecture description

Step 2: Load Appropriate Checklist

Based on tech stack, load the corresponding checklist from references/checklists.md:

• FastAPI: API design, Pydantic models, async patterns, security
• Next.js: SSR/SSG, routing, state management, performance
• General: Universal best practices

Step 3: Build and Execute Gemini Command

Load the appropriate prompt template from references/prompt-templates.md and execute:

For inline prompts:

gemini -m gemini-3-pro-preview -p "{constructed_prompt}" --output-format json

For file content review:

cat {file_path} | gemini -m gemini-3-pro-preview -p "{review_instructions}" --output-format json

For multi-line prompts (heredoc):

gemini -m gemini-3-pro-preview --output-format json -p "$(cat << 'EOF'
{constructed_prompt}
EOF
)"

Step 4: Parse JSON Response

Gemini returns JSON with this structure:

{
  "response": "The actual review content",
  "stats": {
    "models": { ... },
    "tools": { ... }
  }
}

Extract the review content from the response field. Use jq for parsing:

result=$(gemini -m gemini-3-pro-preview -p "..." --output-format json)
echo "$result" | jq -r '.response'

Step 5: Present Results in Two Sections

CRITICAL: Always present results in two clearly separated sections:

Section A: Gemini Raw Response

Present Gemini's complete response without modification. This shows the user exactly what Gemini found.

Section B: Claude's Analysis & Action Plan

Based on Gemini's feedback, Claude provides:

1. Summary (2-3 sentences)

   • Key findings
   • Overall assessment

2. Action Items (prioritized list)

   • 🔴 Critical: Must fix before proceeding
   • 🟡 Important: Should address soon
   • 🟢 Minor: Nice to have improvements

3. Ready-to-Apply Code (when applicable)

   • Provide actual code fixes for each issue
   • Include before/after comparison when helpful

Output Format Template

---
## 📋 Gemini Review Results

### A. Gemini Response (Raw)

{gemini_response_verbatim}

---

### B. Claude's Analysis

#### Summary
{2-3 sentence overview}

#### Action Items

🔴 **Critical**
- Issue: {description}
- Fix: {solution}

🟡 **Important**  
- Issue: {description}
- Fix: {solution}

🟢 **Minor**
- Issue: {description}
- Fix: {solution}

#### Ready-to-Apply Code

**{issue_name}**
```{language}
{fixed_code}

---

## Command Examples

### Plan Review
```bash
gemini -m gemini-3-pro-preview --output-format json -p "$(cat << 'EOF'
[PLAN REVIEW REQUEST]
Review this implementation plan for completeness and potential issues:

{plan_content}

Check for: Logic errors, missing edge cases, architecture flaws, security concerns.
Provide specific, actionable feedback.
EOF
)"

Code Review (piping file)

cat src/api/auth.py | gemini -m gemini-3-pro-preview -p "Review this FastAPI authentication code for security issues, bugs, and best practices. Provide specific line-by-line feedback." --output-format json

Architecture Review

gemini -m gemini-3-pro-preview --output-format json -p "$(cat << 'EOF'
[ARCHITECTURE REVIEW]
System: {system_name}
Tech Stack: FastAPI + Next.js + AWS

{architecture_description}

Evaluate: Scalability, reliability, maintainability, security, cost efficiency.
Suggest improvements with trade-off analysis.
EOF
)"

Environment Compatibility

Environment	Execution Method
Claude Code	Direct bash execution
Claude (Web/Desktop)	Use bash_tool

Command syntax remains identical across environments.

Error Handling

If Gemini returns an error:

1. Display the error message to user
2. Check for common issues:
   • Quota exceeded (60/min or 1000/day limit)
   • Network connectivity
   • Invalid prompt format
3. Suggest retry or alternative approach

Quota Management

Free tier limits:

• 60 requests per minute
• 1000 requests per day

To conserve quota:

• Combine related reviews into single prompts
• Use specific, focused review requests
• Avoid redundant re-reviews

References

• references/checklists.md: Tech stack-specific review checklists
• references/prompt-templates.md: Prompt templates for each review type