Skills¶

CLIver has an LLM-driven skill system that allows the agent to discover and activate specialized capabilities during a chat session. Skills are defined as SKILL.md files with YAML frontmatter.

How Skills Work¶

Unlike traditional configuration-based systems, CLIver's skills are LLM-driven:

The skill builtin tool is always available (core tool)
The LLM calls skill('list') to discover available skills
The LLM calls skill('skill-name') to activate a skill and receive its instructions
The skill content is injected into the conversation as context

This means the LLM decides when to use skills based on the user's request — no manual flags needed.

Creating a Skill¶

A skill is a directory containing a SKILL.md file with YAML frontmatter:

my-skill/
└── SKILL.md

SKILL.md Format¶

---
name: code-review
description: Performs thorough code reviews with security and quality analysis
tags:
  - code
  - review
  - security
---

# Code Review Skill

When reviewing code, follow these steps:

1. Check for security vulnerabilities (SQL injection, XSS, etc.)
2. Review code quality and adherence to best practices
3. Identify potential performance issues
4. Suggest improvements with examples

## Output Format

Provide your review in sections:
- **Security**: List any security concerns
- **Quality**: Code quality observations
- **Performance**: Performance considerations
- **Suggestions**: Actionable improvements

Frontmatter Fields¶

Field	Required	Description
`name`	Yes	Unique skill identifier
`description`	Yes	Brief description (shown in skill list)
`tags`	No	Keywords for discovery

Skill Discovery Paths¶

Skills are loaded from two locations (project-local takes priority):

Project: .cliver/skills/ in the current working directory
Global: ~/.config/cliver/skills/

Each skill is a subdirectory containing a SKILL.md file.

Workflow Integration¶

Skills can be activated within workflow LLM steps using the skills field:

name: review-workflow
steps:
  - id: review
    type: llm
    name: Code Review
    prompt: "Review the code in {{ inputs.file_path }}"
    skills:
      - code-review
    outputs: [review_result]

When a step declares skills, the skill content is deterministically injected into the LLM context (not probabilistic activation).

Example Skills¶

Greeting Skill¶

---
name: greeting
description: Greets users in a friendly and personalized way
tags:
  - greeting
  - welcome
---

# Greeting Skill

Greet the user warmly. If you know their name (from identity or memory), use it.
Mention the current time of day if possible.

Research Skill¶

---
name: research
description: Structured research with web search and analysis
tags:
  - research
  - web
  - analysis
---

# Research Skill

When conducting research:

1. Use `web_search` to find relevant sources
2. Use `web_fetch` to read promising results
3. Synthesize findings into a structured summary
4. Cite sources with URLs