The prompt contains the instructions you write for an agent: its role, responsibilities, tone, and how it should respond to users.

## Edit your instructions

Open the agent's **Prompt** settings and write in the editor.

Your instructions are what the agent reads whenever it responds. Product documentation is accessed through [Knowledge](/agent-knowledge), while the full instructions in a [skill](/skills) load when the agent requests that skill. These let you provide product context and task-specific instructions without putting everything in the prompt.

## Reference tools, actions, and skills

Type `@` in your instructions to open the reference picker. It has three groups:

- **Tools:** the tools associated with this agent's configuration, including built-in tools.
- **Actions:** the API actions configured for this agent.
- **Skills:** the reusable instructions configured for this agent.

A reference does not grant access to a capability. If you remove a grant that an instruction still references, the editor marks the reference as stale — the chip stays visible, dashed and tinted, with its cause in the tooltip, and it is still kept as written. Update or remove it before saving; Teaflask validates references against the agent's configuration and refuses a save that names something the configuration does not give the agent.

See [Mentions](/mentions) for the reference syntax.

## Test and publish

Save your changes and try the draft in the [Playground](/playground). Editing or saving the prompt does not replace the published agent configuration. [Publish a revision](/revisions) when you want users to receive the updated instructions.

## Continue reading

- [Skills](/skills): reusable instructions that load on demand.
- [Agent knowledge](/agent-knowledge): the documentation the agent can access.
- [Model](/model): model selection and conversation-context settings.