Skip to main content

Agent Skills

A skill is a reusable operating procedure that an agent loads only when a task matches it. The minimum package is a directory containing SKILL.md; scripts, references, assets, and interface metadata are optional.

What Belongs in a Skill​

  • a workflow repeated across sessions or repositories;
  • domain rules that materially change how work is performed;
  • a reliable sequence with clear inputs, outputs, checks, and stopping conditions;
  • scripts or templates that are safer to reuse than regenerate.

Do not create a skill for a one-off preference, a fact that belongs in normal documentation, or a thin wrapper around one obvious command.

Design Rules​

  1. Make the description a precise trigger: what it does, when to use it, and important exclusions.
  2. Keep SKILL.md as the route map and operating contract.
  3. Move detailed references and deterministic mechanics into focused files or scripts.
  4. Keep references shallow; an agent should not chase a maze before acting.
  5. State authority boundaries and validation proportional to risk.
  6. Test the skill forward on a real task, not only by checking its folder shape.

This is progressive disclosure: metadata is always visible, full instructions load on activation, and supporting resources load only when required.

Relationship to Other Layers​

  • Instructions say how an agent should generally behave in a scope.
  • A skill teaches a repeatable method.
  • MCP supplies a standardized connection to live tools or context.
  • The harness runs the loop and observes whether the method worked.

For this website, the site-management skill is the general entry point for requests such as “put this on my website.” The content itself still lives in ordinary Markdown and remains understandable without the skill.