What it is
Skills are portable folders holding instructions (SKILL.md), and often scripts and assets too, that teach an agent a specific procedure: "how to make slides", "how to edit video", "how to review code against our checklist". The agent loads a skill on demand, without bloating its permanent context.
Where it came from
- October 2025 — Anthropic announces Agent Skills: modular, versionable skills for Claude (Code, the app, the API).
- The idea echoes sub-agents and AGENTS.md: move a repeated procedure into an artifact instead of keeping it in the prompt.
- Skill marketplaces and collections started appearing (skill-creator, domain-specific bundles).
Why it took off
- It solves the "context is full" problem: a skill loads only when needed (progressive disclosure).
- Portability: one skill works across different projects and surfaces.
- A repeated manual pattern becomes a skill: the team stops explaining the same thing to the agent every time.
How to use it today
- Notice you're repeating the same procedure to an agent — package it as a skill (SKILL.md plus scripts if needed).
- Keep skills narrow and well described — the agent will decide when to load them.
- Before installing someone else's skill — read SKILL.md and the sources (supply-chain audit) and restrict its permissions.
What to watch out for
- A skill with executable scripts is third-party code: the same risk as any package or MCP server.
- A poorly described skill means the agent won't activate it in time (or will activate it in the wrong place).
- Version them: a stale skill can drag in an outdated procedure.