ask-gemini

Ask Gemini Skill

Collaborate with Google Gemini for research, code review, feedback, validation, or second opinions.

Purpose

This skill enables model collaboration where Claude can delegate work to Gemini to get:

• Alternative perspectives on architectural decisions
• Code review from a different model
• Research validation and enhancement
• Edge case identification
• Performance and security feedback

When to Use

Use this skill when:

• You need a second opinion on important decisions
• You want to validate code quality before committing
• You're researching complex topics and want multiple perspectives
• You need to identify blind spots or edge cases
• You want to enhance specifications or documentation
• A subagent needs expert consultation

Parameters

The user will invoke this with:

/ask-gemini "prompt" "context"

Or from a subagent:

Skill: ask-gemini
Args: prompt="Review this code" context="src/file.js"

Required:

• prompt - The question or task for Gemini (what you want Gemini to do)

Optional:

• context - Files, code, or additional information for Gemini to consider
  • Can be file paths (e.g., src/parser.js)
  • Can be directory paths (e.g., specs/)
  • Can be multiple files (space-separated)
  • Can be text/code snippets

Execution Instructions

Step 1: Parse Input

Extract the prompt and context from user input:

• If args provided, use those
• Otherwise parse from natural language

Step 2: Gather Context

If context includes file paths:

1. Use Read tool to read the files
2. If directory, use Glob to find relevant files
3. Limit to reasonable size (don't send 100+ files)
4. Summarize if files are very large

Step 3: Craft Enhanced Prompt for Gemini

Create a rich prompt that includes:

1. The user's original prompt
2. Relevant context from files
3. Project information (if helpful)
4. Specific questions or focus areas

Example Enhanced Prompt:

[User's Question]
Review this specification for completeness and edge cases.

[Context]
File: specs/001-code-wiki/spec.md
```markdown
[file contents here]

[Additional Context]

• This is part of a documentation generation system
• Target audience: developers
• Performance requirements: handle repos up to 100k files

[Specific Focus] Please identify:

1. Missing edge cases
2. Unclear requirements
3. Potential performance issues
4. Security considerations
5. Suggested improvements

### Step 4: Call Gemini via CLI

Use the Gemini CLI directly via Bash tool:

```bash
# Create a temporary file with the enhanced prompt
echo "[enhanced prompt with context]" > /tmp/gemini-prompt.txt

# Call Gemini CLI
gemini --model gemini-3-pro-preview < /tmp/gemini-prompt.txt

# Or with streaming:
gemini --model gemini-3-pro-preview --stream < /tmp/gemini-prompt.txt

Alternative: Direct pipe

echo "[enhanced prompt]" | gemini --model gemini-3-pro-preview

Model Selection:

• Default: gemini-3-pro-preview (most capable, deep reasoning)
• For quick questions: gemini-2.5-flash (ultra-fast)
• For code generation: gemini-3-pro-preview
• For latest features: gemini-2.0-flash (audio/video support)

CLI Flags:

• --model - Specify Gemini model
• --stream - Stream response (optional)
• --temperature - Control randomness (0-1)
• --json - Output JSON format (if needed)

Step 5: Process Response

1. Capture Gemini's CLI output
2. Format it nicely with markdown
3. Extract key points (if response is long)
4. Save conversation context for follow-ups (if needed)

Step 6: Present to User

Format output as:

## Gemini's Analysis

[Gemini's response, formatted nicely]

---

## Summary (if response is long)

**Key Points:**
- [bullet point 1]
- [bullet point 2]
- [bullet point 3]

**Recommendations:**
- [action item 1]
- [action item 2]

---

## Next Steps

Would you like me to:
1. [Specific action based on Gemini's feedback]
2. [Another specific action]
3. Ask Gemini a follow-up question

Step 7: Offer Actions

Based on Gemini's response, offer to:

• Implement suggestions
• Update files with feedback
• Create tasks for follow-up work
• Ask Gemini follow-up questions (preserve continuation_id)

Examples

Example 1: Code Review

Input:

/ask-gemini "Review for bugs and security issues" src/parser.js

What Claude Does:

1. Reads src/parser.js
2. Creates prompt: "Review this code for bugs, security vulnerabilities, and best practices..."
3. Calls Gemini with code context
4. Returns formatted review
5. Offers to fix identified issues

Example 2: Spec Enhancement

Input:

/ask-gemini "Identify missing requirements and edge cases" specs/spec.md

What Claude Does:

1. Reads specs/spec.md
2. Analyzes current spec structure
3. Asks Gemini for gaps, edge cases, unclear areas
4. Returns comprehensive feedback
5. Offers to update spec with improvements

Example 3: Architecture Decision

Input:

/ask-gemini "Compare REST vs GraphQL for our documentation API"

What Claude Does:

1. Gathers project context
2. Creates structured comparison prompt
3. Gets Gemini's analysis with pros/cons
4. Returns detailed comparison
5. Offers to create decision document

Example 4: Research Validation

Input:

/ask-gemini "Validate this research and suggest additional sources" research-notes.md

What Claude Does:

1. Reads research notes
2. Asks Gemini to validate facts and add perspectives
3. Gets additional sources and corrections
4. Returns enhanced research
5. Offers to update notes

Advanced Usage

Follow-up Questions

For follow-up questions, Claude can maintain context by:

1. Saving previous conversation in a temporary file
2. Including it in subsequent prompts to Gemini
3. Building on previous responses

User: /ask-gemini "What about performance?"
Claude: [Includes previous Q&A context in new prompt to Gemini]

Multi-file Context

/ask-gemini "Compare these implementations" src/v1.js src/v2.js

Directory Analysis

/ask-gemini "Review architecture patterns" src/

Tips for Effective Use

1. Be Specific: "Review for security issues" > "Review this"
2. Provide Context: Include relevant files, not entire codebase
3. Ask Focused Questions: Narrow scope = better responses
4. Use for Important Decisions: Two perspectives = better outcomes
5. Iterate: Use continuation for follow-up questions

Error Handling

If Gemini CLI fails:

1. Check if Gemini CLI is installed (which gemini or gemini --version)
2. Verify API key is set in environment (usually GOOGLE_API_KEY or GEMINI_API_KEY)
3. Check if model name is correct
4. Try with smaller context (if prompt too large)
5. Check for rate limiting or quota issues
6. Fall back to Claude-only analysis with note to user

Technical Notes

• Uses Gemini CLI directly via Bash tool
• Context maintained via conversation history in prompts
• Context window: Gemini supports 1M tokens (model dependent)
• Model selection: Defaults to gemini-3-pro-preview
• Can be invoked by subagents via Skill tool
• Requires Gemini CLI to be installed and configured

Output Format

Always return structured output:

1. Clear heading
2. Gemini's response (formatted)
3. Summary of key points
4. Actionable next steps
5. Offer to implement or follow up

---

Remember: This skill enables model collaboration for better outcomes. Use it when a second perspective adds value!