Skills are reusable instructions maintained by your organization and shared across agents. Creating a skill, publishing its instructions, and granting an agent access are separate steps.

## Create and publish a skill

Open **Skills** and create a skill with a name and description, then write its instructions. The description tells agents when the skill applies. Names use lowercase letters, numbers, and hyphens, and are permanent: to use a different name, create a new skill.

Save your changes, then publish the skill revision you want agents to use. Saving alone does not update the published instructions. Republishing a skill updates it for every agent that already grants it, without republishing those agents. The published revision is resolved at the start of a turn, so publishing does not replace instructions midway through that turn.

## Grant skills to an agent

Open the agent's **Skills** settings and select the individual skills it may load. A draft skill can be granted, but the agent cannot load it until it has a published revision.

**Skill activation** controls approval for loading the granted skills:

- **Allow:** the agent can load a skill without requesting approval.
- **Ask:** loading a skill pauses for the user's approval.

Save and [publish the agent](/revisions) to make grant and approval changes available to users. Publishing a skill does not grant it to an agent automatically.

## How skills load

The agent receives the names and descriptions of available skills. Their full instructions load on demand when the agent requests a skill; they are not included in every turn like the agent's [prompt](/prompt).

A skill's instructions do not grant tools, actions, or other skills. Those capabilities must also be configured on the agent. See [Mentions](/mentions) for referencing them in instructions.

If a granted skill is deleted or has no published revision, it is skipped and the agent continues without it. Remove stale grants from the agent's settings when a skill is no longer available.

## Continue reading

- [Prompt](/prompt): instructions included whenever the agent responds.
- [Actions](/actions): grant API operations referenced by a skill.