biome

Agent Skills are directories containing a SKILL.md file that packages domain expertise and organizational knowledge for reuse across conversations. They enable Claude to perform specialized tasks by loading instructions, scripts, and resources dynamically.

Key Characteristics:

• Model-invoked: Claude autonomously decides when to use them based on the description
• Progressive disclosure: Only loads content when needed, optimizing token usage
• Composable: Multiple focused skills work better together than one comprehensive skill

1. Create Your Skill Directory

```

my-skill/

├── SKILL.md # Required - core instructions

├── reference.md # Optional - additional documentation

├── examples.md # Optional - usage examples

├── scripts/ # Optional - executable code

└── templates/ # Optional - template files

```

2. Write Your SKILL.md

```yaml

---

name: your-skill-name

description: Brief description of what this Skill does and when to use it

---

# Your Skill Name

Provide clear, step-by-step guidance for Claude.

Show concrete usage examples with inputs and outputs.

Best practices and constraints.

```

3. Install Your Skill

| Location | Scope |

|----------|-------|

| ~/.claude/skills/ | Personal - available across all projects |

| .claude/skills/ | Project - specific to repository |

| Field | Required | Constraints |

|-------|----------|-------------|

| name | Yes | Max 64 chars, lowercase letters, numbers, hyphens only |

| description | Yes | Max 1024 chars, explains function and trigger conditions |

| version | No | Semantic versioning (e.g., 1.0.0) |

| dependencies | No | List of required packages |

Skills implement three levels of information loading:

1. Level 1 (Metadata): Name/description loaded into system prompt at startup
2. Level 2 (Instructions): Full SKILL.md loads when Claude determines relevance
3. Level 3+ (Resources): Additional files load only when referenced

This design keeps context consumption minimal while enabling unbounded skill complexity.

• Focus: Create separate skills for different workflows
• Clarity: Write specific descriptions for proper invocation
• Simplicity: Start with markdown; add code only when needed
• Examples: Include sample inputs and outputs
• Versioning: Track changes systematically
• Incremental Testing: Test after each significant change
• Scale Structure: Split into separate files when SKILL.md grows unwieldy

• Install skills only from trusted sources
• Audit bundled files and dependencies
• Never hardcode sensitive information (API keys, passwords)
• Review instructions directing Claude toward external connections

```

claude-skills-marketplace/

├── skills/ # Community skills

│ ├── creative/ # Creative & design skills

│ ├── development/ # Development & technical skills

│ ├── enterprise/ # Enterprise & communication skills

│ └── document/ # Document generation skills

├── templates/ # Starter templates

│ └── basic/

│ └── SKILL.md

├── docs/ # Documentation

│ ├── CREATING_SKILLS.md

│ ├── API_SKILLS.md

│ └── SKILLS_BEST_PRACTICES.md

└── CONTRIBUTING.md

```

See [CONTRIBUTING.md](CONTRIBUTING.md) for guidelines on submitting skills to the marketplace.

• [Introducing Agent Skills | Anthropic](https://www.anthropic.com/news/skills)
• [How to create custom Skills | Claude Help Center](https://support.claude.com/en/articles/12512198-how-to-create-custom-skills)
• [GitHub - anthropics/skills](https://github.com/anthropics/skills)
• [Equipping agents for the real world | Anthropic Engineering](https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills)

MIT