gemini（Gemini 助手）

Gemini Skill Guide

When to Use This Skill

Primary Use Cases:

1. User explicitly requests gemini CLI execution (Mode 1: Direct CLI)
2. Large-context analysis requiring 1M token window (Mode 2: Gemini Delegation)
3. Repository-scale code reviews and architectural analysis
4. Multimodal analysis (code + images/diagrams)
5. Cross-module dependency tracking

Do NOT use for:

• Simple code explanations (use Claude directly)
• Small context tasks (<100K tokens)
• Tasks not requiring Gemini-specific capabilities

Execution Modes

Mode 1: Direct CLI Execution

When: User explicitly asks to run gemini CLI

Workflow:

1. Gather configuration using structured questions:

- Model selection (gemini-3-pro-preview vs gemini-3-flash-preview) - Approval mode (default, auto_edit, yolo) - Output format (text, json, stream-json)

1. Execute: gemini "<PROMPT>" -m <MODEL> --approval-mode <MODE> --output-format <FORMAT>
2. Inform user: "You can resume with 'gemini --resume latest' anytime"

Resuming:

gemini "<CONTINUATION PROMPT>" --resume latest

Mode 2: Gemini Delegation for Large Context

When: Task requires large context window (>100K tokens), multimodal input, or repository-scale analysis

Workflow:

1. Analyze context and identify what Gemini analysis would help
2. Formulate comprehensive query with all relevant context:

- Problem statement - Current findings - Code snippets or file references - Specific questions

1. Leverage Gemini's 1M token context for comprehensive analysis
2. Execute: gemini "<DETAILED CONTEXT>" --output-format json
3. Synthesize Gemini response into actionable insights
4. Report findings with:

- Clear summary of Gemini's analysis - Specific recommendations or solutions - Additional considerations or caveats - Next steps if applicable

Error Handling

• Verify gemini binary exists before execution: which gemini
• Stop immediately on non-zero exit codes and report to user
• Request direction before retrying failed commands
• Before using high-autonomy flags, confirm with user:

- --yolo (auto-approve all actions) - --approval-mode yolo

Context Management

• Gemini excels at large context (up to 1M tokens)
• For massive contexts (>500K tokens), consider module-by-module summaries
• Use --include-directories to add additional workspace paths
• Leverage session resume for iterative refinement

Configuration Gathering Patterns

Initial Configuration (Mode 1, Step 1)

At the start of Mode 1 execution, gather model, approval mode, and output format:

Standard configuration question:

Select gemini configuration (model, approval mode, and output format):

1. "gemini-3-pro-preview / default / json (Recommended)" - High quality, safe mode
2. "gemini-3-pro-preview / auto_edit / json" - High quality, auto-approve edits
3. "gemini-3-flash-preview / default / json" - Fast, safe mode
4. "gemini-3-flash-preview / auto_edit / json" - Fast, auto-approve edits
5. "gemini-3-pro-preview / yolo / json" - High quality, fully autonomous
6. "gemini-3-flash-preview / yolo / json" - Fast, fully autonomous
7. "Custom" - User will specify model, approval, and format separately

High-Impact Flags Confirmation

Before executing with --yolo or autonomous flags:

Ready to execute with these flags: [LIST FLAGS]. Proceed?

1. "Execute now" - Run as configured
2. "Modify configuration" - Change settings
3. "Cancel" - Abort

Post-Execution Follow-up

After gemini command completes:

Gemini completed. [SUMMARY]. Next steps?

1. "Resume with additional prompt" - Continue session
2. "Analyze results" - Review output
3. "Complete" - Finished
4. "Retry with different config" - Adjust settings

Error Recovery

When command fails or has warnings:

Error: [SPECIFIC ERROR]. How to proceed?

1. "Resume with adjustments" - Fix and continue
2. "Retry with different config" - Change model/approval/format
3. "Accept partial results" - Use what worked
4. "Invoke heal-skill" - Fix outdated SKILL.md

Running a Task

1. Gather configuration using the patterns above
2. Assemble the command with the appropriate options:

- -m, --model <MODEL> (gemini-3-pro-preview or gemini-3-flash-preview) - --approval-mode <MODE> (default, auto_edit, yolo) - -o, --output-format <FORMAT> (text, json, stream-json) - --include-directories <DIRS> (additional workspace paths) - -r, --resume <SESSION> (latest or session index)

1. Run the command, capture output, and summarize for the user
2. After Gemini completes, inform the user: "You can resume this Gemini session at any time by saying 'gemini resume' or asking me to continue."

Quick Reference

Model Selection

Session Management

# List available sessions
gemini --list-sessions

# Resume most recent session
gemini --resume latest

# Resume specific session by index
gemini --resume 5

# Delete a session
gemini --delete-session 3

MCP Server Integration

# List configured MCP servers
gemini mcp list

# Add an MCP server
gemini mcp add <name> <command> [args...]

# Remove an MCP server
gemini mcp remove <name>

# Limit to specific MCP servers
gemini "<PROMPT>" --allowed-mcp-server-names mcp-skillset

Extension System

# List installed extensions
gemini extensions list

# Install extension from git or path
gemini extensions install <source> [--auto-update] [--pre-release]

# Update extensions
gemini extensions update [--all]

# Disable/enable extension
gemini extensions disable <name>
gemini extensions enable <name>

Following Up

• After every gemini command, offer follow-up options
• When resuming, use: gemini "<new prompt>" --resume latest
• The resumed session automatically uses the same context from the original
• Restate the chosen model and approval mode when proposing follow-up actions

Error Handling Guidelines

• Stop and report failures whenever gemini --version or a gemini command exits non-zero; request direction before retrying
• Confirm high-impact flags before execution
• When output includes warnings or partial results, offer error recovery options

Comparison with Codex Skill

Integration Patterns

With Other CLI Agents

# Use gemini for large context, codex for GPT-specific
gemini "Analyze entire codebase architecture" --output-format json
codex exec "Implement specific feature based on analysis" --full-auto

# Chain with research CLI
research docs -t "framework API" --format json | gemini "Apply this to codebase"

Multimodal Analysis

# Gemini can process images inline with code
gemini "Analyze this architecture diagram and compare with implementation" \
  --include-directories ./docs ./src

Heal-Skill Integration

When gemini CLI API changes are detected (command failures, unexpected output formats, or deprecated flags):

1. Detection: Notice command failures or API mismatches during execution
2. Trigger: Flag skill for healing via /heal-skill gemini
3. Analysis: Healing agent analyzes current CLI with gemini --help and gemini --version
4. Update: Updates skill documentation to match current API
5. Validation: Re-validates agent configuration for compatibility
6. Model Verification: Ensures only gemini-3 models are referenced
7. Context Window Verification: Confirms 1M token context window specifications

Common Changes to Monitor:

• New or deprecated command flags
• Changes to approval modes or output formats
• Model availability updates (gemini-3 family)
• MCP server integration API changes
• Session management API modifications
• Extension system updates
• Context window specification changes