Anthropic Agent Skills

One-liner: Anthropic's spec for composable, file-based agent capabilities — each Skill is a folder containing a SKILL.md with YAML frontmatter plus any supporting scripts or references, loaded on-demand via progressive disclosure.

Anatomy of a Skill

• A directory at a well-known path (e.g. .claude/skills/<name>/).
• A required SKILL.md file with YAML frontmatter:
• name — short identifier used to invoke the skill
• description — when to use it; this is what the model reads to decide relevance
• optional fields for tool allowlists, version, model hints
• Optional sibling files: reference.md, scripts/, examples — only loaded if the skill is selected.

Progressive disclosure

The agent sees only the frontmatter index at first. The full SKILL.md body is loaded when the model selects the skill; supporting files are loaded only when the skill body references them. Keeps the context budget small until a skill is actually needed.

Why it matters

• The same shape underpins this wiki's own automation skills (wiki-pipeline, wiki-add, wiki-query, wiki-review). A personal-wiki system is already an Agent Skills consumer.
• Skills are portable across Claude Code, Claude API, and third-party tools that adopt the spec.
• Composes well with MCP: a Skill can describe when and how to use a tool, while MCP provides the tool itself.

Open questions

• Skill discovery / marketplace conventions still emerging.
• Versioning and update propagation across machines not yet standardised.
• Permission model for tool allowlists inside frontmatter is evolving.

Changelog

• 2026-05-10 — Page created from https://platform.claude.com/docs/agent-skills (Type B, confidence 90)