Skip to main content

Skills

A skill is a named set of Markdown instructions the model loads only when it needs them. Each skill has a slug ($docs-style), a display name, a short description, and the full instructions.

Lazy loading​

The description is the manifest: when tools are enabled for a turn, the built-in load_skill tool lists every enabled skill's slug, name, and description, and the model fetches the full instructions with one call. Long instructions therefore never occupy context until they are used. Mentioning $slug in a message nudges the model toward loading that skill.

Lifecycle​

Skills can be disabled without deleting them; a disabled skill leaves the manifest and refuses to load. Updates bump the version and archive the prior revision, with rollback restoring any revision as a new version. The interchange form is a SKILL.md document — frontmatter carrying the name, slug, and description, followed by the Markdown instructions verbatim — which export downloads and import accepts alongside the JSON envelope. Names, descriptions, and instructions are encrypted at rest; the slug stays plain.

Importing from a store​

Skills can be pulled straight from a remote source under Settings → Skills → From URL. Accepted forms: a skills.sh listing link, a GitHub repository or tree/blob URL, an owner/repo/skill shorthand (resolved against the conventional skills/<name>/SKILL.md layout), or a direct link to any SKILL.md file. The fetch runs through the same pinned egress guard as tool servers — public addresses only, size capped — and the document lands as your own editable skill, so a remote source never updates itself silently.

Bundled files​

A skill can carry companion files in the SKILL.md folder layout — reference documents, templates, and script sources the instructions point at. load_skill lists the bundle at the end of the instructions, and the model reads a file on demand with the read_skill_file tool, so a large reference never occupies context until it is actually needed.

Files travel everywhere the skill does: Settings → Skills → From folder uploads a local SKILL.md folder in one step, a remote import pulls the text files sitting next to the fetched SKILL.md, the JSON export carries the bundle, and the edit modal manages files one by one. Paths are relative and can never escape the skill; contents are text-only, encrypted at rest, and bounded (32 files, 200 KB each, 1 MB per skill).

Taught skills​

Skills whose slug begins with taught- were created by demonstrating a task on a Work Computer screen: the recording is compiled deterministically into a natural-language playbook — anchored click targets, redacted secrets, and an allowed scope derived from the sites the demonstration visited — and saved as an ordinary skill, so it gets versioning, editing, and sharing like any other. Computer-enabled Work runs load the owner's enabled taught skills automatically. After a run finishes, a one-click worked/failed review on the run's skill chips appends a dated line to the skill's Track record section; each review is a normal skill version, so the history is editable and restorable here like any other change.

Required approvals​

A skill can carry an approval demand alongside its instructions. In the skill editor, Always ask before turns it on and a checklist picks the tools it covers: run_command, computer_act, delete_file, move_file, message_agent, and write_file. Leaving every box unchecked covers all six.

While the skill is loaded — bound to the profile or persona for the turn, or loaded mid-turn with load_skill — those tools pause for the user before they run, no matter what the run's own approval setting says, and the request names the skill that is asking. A saved "Always allow" rule can never approve them: prose in the instructions is advice the model may paraphrase past, while this is enforced at the gate. A one-time "Allow once" still works, because that is a live decision rather than a stored rule. See action approvals.

The field travels with the skill: it round-trips through export and import in the JSON envelope. The SKILL.md interchange form carries instructions only, so a markdown-only import lands on the default (inherit) and the demand has to be set again.

A taught skill whose demonstration redacted a secret is saved with this turned on for computer_act: replaying it goes back through territory the user must see, and the prose boundary alone was never enough.

Sharing and profiles​

Skills are private by default and shareable through the common resource-grant model. An assistant profile can bind a subset of skills, and sessions using that profile only see those skills in the manifest.

Example​

Create a skill under Settings → Skills (or pick one of the starter templates there and edit it):

  • Slug: cite-sources

  • Name: Cite sources

  • Description: How to cite evidence: numbered references with URLs, no unreferenced claims. (this line is what the model sees in the manifest — make it say when to use the skill)

  • Instructions:

    # Citing sources

    - Every factual claim gets a numbered reference like [1].
    - End with a References section: number, title, URL.
    - If no source exists, say so instead of inventing one.

Enable the tools toggle in the composer and ask:

Summarize the latest results and $cite-sources.

Typing $ autocompletes your skill slugs. The model sees cite-sources in its manifest, calls load_skill to fetch the instructions above, and formats the answer accordingly. Disable the skill on its card and it disappears from the manifest without being deleted.