Skip to main content

What is a skill?

A skill is a folder of instructions that teaches the agent how to do one specific job, such as “draft release notes from a changelog” or “produce our standard weekly sales report”. It contains a SKILL.md file and, optionally, scripts, templates, or reference files the instructions refer to. Skills keep your agent’s base instructions short. Instead of putting every procedure in the agent’s instructions, you install skills in its sandbox. The agent sees each skill’s name and description, and loads the full instructions only when a task calls for that skill.
SKILL.md starts with YAML front matter giving a name and a description. The name uses only letters, digits, _ and -, since it names the folder the skill is staged in; an upload with any other name is refused. Write the description so the model can tell when to use the skill:
Skills need a managed sandbox; they are installed as files inside it.

Versions

Every upload of a skill is a numbered version. The skill also has a default version. Uploading a new version does not change the default unless you ask it to, so you can test a new version before switching everyone to it. When you reference a skill, you can pin a specific version. Pinning makes runs reproducible: a session always gets exactly the instructions you tested.

Example: upload, version, and use in a template

Examples on this page assume client is an OpenAI client configured as in the quickstart. The upload is a zip file containing the skill folder. The zip can be at most 32 MiB, with at most 2,048 entries, each at most 32 MiB, and at most 128 MiB once unpacked.

Install a skill in a session

Add it to the session’s environment, or to a template the session uses:

Environment templates

An environment template is a saved sandbox setup that many sessions can share: staged files, skills, packages, setup commands, environment variables, and network policy. Create it once, then pass its ID:
Settings on the session override the template’s, except that a session cannot loosen the template’s network policy. Do not delete a template or skill while sessions that depend on it are still being created.

Keep it safe

  • Review skills like code. A skill’s instructions direct what the agent does, and its scripts run in the sandbox. Read third-party skills before installing them.
  • No secrets in skills or templates. Anything in them is readable by the agent. Use vault credentials for tokens.
  • Self-hosted environments cannot install skills. Put the files the agent needs in the workspace folder instead.
See also: execution environments, runnable resource examples, and the skills reference.