Skip to content

gemini_cli_skills.md

Latest commit

 

History

History
60 lines (42 loc) · 2.33 KB

gemini_cli_skills.md

File metadata and controls

60 lines (42 loc) · 2.33 KB
Raw

Loading New Skills in Gemini CLI

The Gemini CLI is a dynamic, agentic interface that automatically discovers capabilities from the local workspace. To extend the agent's intelligence with a new skill, follow these integration steps.

1. Directory Structure

The CLI specifically looks for skills in the .gemini/skills/ directory. Every new skill MUST follow this structure to be discoverable:

.gemini/skills/my-new-skill/
├── SKILL.md              <-- This is the intelligence contract
└── scripts/              <-- This is the execution logic
    ├── my_tool.py
    └── test_my_tool.py   <-- Optional: recognized by 'test-all'

2. Defining the Intelligence Contract (SKILL.md)

The SKILL.md file serves as the documentation the Gemini agent reads to understand when and how to use your skill. It MUST include a YAML frontmatter:

---
name: my-new-skill
description: Describe EXACTLY what this skill does and when the agent should pick it.
---

# My New Skill Documentation
Add usage examples and command line patterns here.

The agent uses the description field to map a user's natural language request to your scripts.

3. Registering the Skill

The Gemini CLI does not require a manual registry. Discovery happens at runtime:

  1. Direct Command: You can force the agent to use it: 
    gemini "Use the my-new-skill to perform [ACTION]"
  2. Autonomous Trigger: If your SKILL.md description is clear, the agent will pick it automatically for relevant tasks.

4. Restoring Skills from the Vault

Many skills in the Croissant Toolkit are provided in an encrypted Vault (.gemini/vault/*.zip) for security. These skills are NOT available to the CLI until they are "Unvaulted":

# Decrypt the skill into the .gemini/skills/ directory
python3 .gemini/skills/odrl-expert/scripts/odrl_client.py unvault-skill "skill-name"

Once unvaulted, the .gemini/skills/<skill-name>/SKILL.md becomes visible, and the agent automatically "learns" its capabilities.

5. Verifying Installation

To ensure the CLI has correctly loaded the skill, you can ask for a system check:

gemini "Run all tests and verify my-new-skill is healthy"

This triggers the test-all skill which discovers and executes any test_*.py files in your new skill's scripts/ folder.