Skills are packaged expertise an AI agent can load when a task needs specific know-how. Unlike tools (which do things), skills teach the agent how to approach a class of work — policies, checklists, writing standards, domain procedures.
In modern agent frameworks, skills often live as Markdown files the runtime injects into context only when relevant. That keeps the default prompt lean while still giving the agent deep guidance on demand.
Skills vs. tools
| Skills | Tools | |
|---|---|---|
| Purpose | What the agent knows / how it should work | What the agent can execute |
| Form | Instructions, playbooks, examples | APIs, functions, code |
| Loaded | When the task matches | When the model decides to act |
| Example | Refund policy checklist | issue_refund API call |
A support agent might load a "billing dispute" skill, then call tools to look up the invoice and draft a reply. The skill shapes judgment; the tools change state.
Why skills matter in a blueprint
Deciding what belongs in skills versus tools versus the system prompt is an architecture choice. Overloading the base instructions with every procedure burns context. Putting judgment rules into opaque code makes them hard to review. Skills keep expertise readable and modular — founders and operators can edit them without touching the runtime.
Related reading
See Tool Use for actions, and Prompt Engineering for how instructions shape behavior.