Gemini Delegation Skill

Delegate specific tasks to Gemini CLI where it has clear advantages. Track token savings after each delegation.

Before Delegating: Check Available Tools

Always check what MCP tools are available:

bash

# List configured MCP servers
gemini mcp list

# Inside interactive session, run:
/mcp list

When to Use Gemini CLI

Delegate these use cases:

1. Large Context Ingestion (2M+ tokens)

Entire repositories, full codebases, multi-file analysis.

bash

gemini -m gemini-3-pro "Analyze the entire codebase for coupling issues"

2. Screenshot & Video Processing

UI screenshots, screen recordings, visual analysis.

bash

gemini -m gemini-3-pro "Analyze this screenshot and describe the layout" < screenshot.png

3. Web Browsing & Fact-checking

Current docs, latest versions, real-time research.

bash

gemini -m gemini-3-pro -e web_search "What are the latest React 19 features?"

4. Browser Automation (via Playwright MCP)

Navigate websites, take screenshots, fill forms, interact with pages.

bash

gemini -m gemini-3-pro "Navigate to https://example.com and take a screenshot"
gemini -m gemini-3-pro "Go to login page, fill username and password, then submit"

5. Bulk Refactoring (High-volume, Low-complexity)

Large-scale find-replace, consistency updates, formatting.

bash

gemini -m gemini-3-pro "Convert all let to const where appropriate across all .ts files"

When to Use Claude Code Instead

Claude handles these better:

•Complex reasoning & autonomous planning - multi-step problem solving
•Big documentation writing - comprehensive docs, architecture docs, guides
•Deep codebase understanding - context-aware modifications
•File operations - Read, Edit, Write tools are faster
•Security-sensitive tasks - safer within Claude's guardrails

Default Model

gemini-3-pro - Default for all delegations (best reasoning capability)

Use gemini-3-flash for faster responses when speed is more important than quality.

Workflow

Step 1: Validate

Is this one of the approved use cases? If NO → handle in Claude directly.

Step 2: Check Tools (if needed)

bash

gemini mcp list

Step 3: Execute

bash

gemini -m gemini-3-pro "your prompt"

Step 4: Report Token Savings

After EVERY Gemini delegation, show this summary:

code

---
**Token Savings Report**
Task: [brief description]
Type: [large-context | screenshot | web-research | bulk-refactor]
Gemini tokens used: ~X,XXX
Estimated Claude tokens saved: ~X,XXX
---

Estimation guide:

•Large context ingestion: 10,000-50,000 tokens saved
•Screenshot processing: 2,000-5,000 tokens saved
•Web research: 3,000-8,000 tokens saved
•Browser automation: 2,000-10,000 tokens saved
•Bulk refactoring: 5,000-20,000 tokens saved (depends on file count)

Quick Reference

bash

# Large context
gemini -m gemini-3-pro "Analyze entire src/ directory..."

# Screenshot
gemini -m gemini-3-pro "Describe this UI" < image.png

# Web research
gemini -m gemini-3-pro -e web_search "Research topic..."

# Browser automation
gemini -m gemini-3-pro "Navigate to URL and take screenshot"

# Bulk refactor
gemini -m gemini-3-pro "Refactor all files to..."

# Check available tools
gemini mcp list

Error Handling

If Gemini CLI fails:

•Check availability: command -v gemini
•Inform user of the error
•Fall back to Claude if appropriate

References

•references/cli-reference.md - Gemini CLI flags and options
•references/mcp-tools-reference.md - MCP servers, Playwright tools, extensions