Gemini Collaboration Skill

This skill enables collaboration between Claude and Gemini CLI for enhanced AI-assisted development workflows.

Prerequisites

Gemini CLI must be installed and authenticated:

• Command: gemini available in PATH
• Version: 0.22.5 or later
• Authentication: Already configured

Operational Modes

Mode 1: Reviewer

Use when the user wants Gemini to review Claude's work.

Triggers: "gemini review", "have gemini check", "get gemini's opinion"

Procedure:

1. Complete the plan, code, or content first
2. Execute the review script:

   $content = @"
   <plan or code content here>
   "@
   $content | powershell -File "skills/gemini-collab/scripts/gemini-review.ps1"
   

3. Present Gemini's feedback to the user
4. Offer to incorporate valid suggestions

Direct Bash invocation:

echo "Review this plan critically. Identify issues, missing considerations, and suggest improvements: <content>" | gemini -o text

Mode 2: Web Searcher

Use when current/recent information is needed that may be beyond Claude's knowledge.

Triggers: "gemini search", "ask gemini about recent", "gemini web search"

Procedure:

1. Identify the information need
2. Execute the search script:

   powershell -File "skills/gemini-collab/scripts/gemini-search.ps1" -Query "search query here"
   

3. Integrate findings into the response

Direct Bash invocation:

gemini "Search for and summarize: <query>" -o text

Mode 3: Co-Builder (Parallel Build)

Use when the user wants both AIs to build the same project for comparison.

Triggers: "both build", "compare builds", "gemini build too", "parallel build"

Procedure:

1. Create the implementation plan
2. Create two directories:
   • claude-build/ - Claude's implementation
   • gemini-build/ - Gemini's implementation
3. Start Gemini build in background:

   powershell -File "skills/gemini-collab/scripts/gemini-build.ps1" -PlanFile "plan.md" -TargetDir "gemini-build" -ApprovalMode "yolo"
   

4. Claude builds in claude-build/ simultaneously
5. Compare results when both complete

Approval Mode Options:

• default - Gemini asks for approval (safer, requires separate terminal)
• yolo - Auto-approve all actions (faster, autonomous)

Direct Bash invocation for Gemini build:

gemini -y "You are building a project. Read the plan and implement it in the gemini-build directory. Plan: <content>" --include-directories ./gemini-build -o text

Mode 4: Collaborative Planner

Use when the user wants both AIs to contribute to planning.

Triggers: "plan together", "collaborative plan", "dual AI planning"

Procedure:

1. Claude drafts initial plan
2. Send to Gemini for review and additions:

   echo "Review this plan. Add missing considerations, suggest alternatives, identify risks: <plan>" | gemini -o text
   

3. Claude refines based on Gemini's input
4. Optionally iterate for convergence
5. Present final collaborative plan to user

Script Reference

gemini-review.ps1

Reviews content and returns structured feedback.

• Input: Content via stdin or -Content parameter
• Output: Gemini's review text

gemini-search.ps1

Performs web search via Gemini.

• Parameter: -Query - The search query
• Output: Search results and summary

gemini-build.ps1

Executes autonomous build in specified directory.

• Parameter: -PlanFile - Path to plan file
• Parameter: -TargetDir - Build directory
• Parameter: -ApprovalMode - default or yolo
• Output: Build status and logs

Best Practices

1. Context Management: Summarize large content before sending to Gemini to avoid token limits
2. Error Handling: Check Gemini CLI availability before invoking
3. Timeout Handling: Long operations should use background execution
4. Output Parsing: Use -o json for structured data when parsing is needed
5. Working Directory: Always use absolute paths or ensure correct working directory

Error Handling

If Gemini CLI is not available:

$geminiPath = Get-Command gemini -ErrorAction SilentlyContinue
if (-not $geminiPath) {
    Write-Error "Gemini CLI not found. Install with: npm install -g @anthropic/gemini-cli"
}

Configuration

The skill respects user preferences for:

• Approval mode (configurable per invocation)
• Output verbosity
• Timeout settings (default: 120 seconds)