Agent Skills Overview (Docs)

What it does

• Defines what a Skill is at the mechanism level: Skills run in Claude's VM environment with filesystem access — directories containing instructions, executable code, and reference materials, organized like an onboarding guide for a new team member
• Explains progressive disclosure concretely: Claude loads information in stages as needed rather than consuming context upfront, with three content levels loaded at different times
• Level 1 is always-loaded metadata — the YAML frontmatter (name, description) that drives discovery, shown with a pdf-processing example
• Points to the Agent Skills Cookbook for authoring your own, and covers SKILL.md structure, bundling, and how the agent discovers and loads skills

When to use it

The authoritative page on how Agent Skills actually work — read it before writing your first SKILL.md so the frontmatter, directory layout, and loading levels are designed rather than guessed.

Getting started

No install — documentation. Author skills as directories with a SKILL.md; the page's pdf-processing frontmatter example is the minimal template to copy.

Things to know

Progressive disclosure only pays off if the metadata level is written for discovery — a vague description means the skill never gets selected, no matter how good its code is.