Agent Skills

Overview

Agent Skills is an open standard from Anthropic for packaging reusable capabilities that AI agents can discover and load dynamically. A skill is a directory containing a SKILL.md file with YAML frontmatter metadata and Markdown instructions. Adopted by Claude, VS Code Copilot, OpenAI Codex, Cursor, and 25+ platforms.

Skill Structure

my-skill/
  SKILL.md          # Required — metadata + instructions
  scripts/          # Optional — executable scripts
  references/       # Optional — reference documents
  assets/           # Optional — supporting files

SKILL.md Format

---
name: my-skill
description: "Use when working with My Framework to follow best practices."
license: MIT
metadata:
  displayName: "My Skill"
  category: frameworks
  author: your-name
compatibility:
  - claude
  - copilot
allowed-tools:
  - Bash
  - Read
  - Edit
---

# My Skill

## Overview
Instructions for the agent when this skill is activated...

## Best Practices
- Concrete guidance the agent should follow
- Code examples demonstrating correct usage

Required Frontmatter

Field	Type	Rules
name	string	Must match directory name. Lowercase, hyphens, digits only. Max 64 chars.
description	string	What the skill does and when to trigger it.

Optional Frontmatter

Field	Type	Description
license	string	SPDX identifier (e.g., MIT, Apache-2.0)
metadata	object	Arbitrary key-value pairs (category, author, tags)
compatibility	list	Target platforms
allowed-tools	list	Tools the skill may use

Description Best Practices

The description field is critical for discovery — agents use it to decide when to activate the skill:

• Start with a trigger phrase: "Use when working with X…"
• Be specific: "Use when configuring ASP.NET Core middleware" not "Web development skill"
• Keep it under 200 characters

Content Guidelines

• Keep the main SKILL.md under 500 lines (~5000 tokens)
• Use concrete code examples, not abstract descriptions
• Include "do this, not that" guidance
• Structure with clear headings for scanability

Validation

# Install the validator
pip install skills-ref

# Validate a skill
agentskills validate path/to/my-skill

# Read properties as JSON
agentskills read-properties path/to/my-skill

# Generate agent prompt XML
agentskills to-prompt path/to/my-skill

Marketplace Publishing

Skills can be published to marketplace repositories:

# Add skills from a marketplace
npx skills add owner/repo

# List available skills
npx skills list owner/repo

The marketplace catalog is defined in .claude-plugin/marketplace.json.

Best Practices

• One skill per concept — keep skills focused on a single framework, tool, or pattern.
• Use trigger phrases in descriptions so agents activate the skill at the right time.
• Validate with agentskills validate before publishing.
• Pin the skill name to the directory name — the validator enforces this.
• Include runnable code examples that agents can adapt, not just documentation.
• Keep instructions actionable: "Do X when Y" rather than "X is a feature of Y."