Instructions
The prompt that defines an agent's role, behavior, and constraints.
Admins only. Members can run and observe; only admins can edit.
Instructions are the prompt that defines how an agent thinks and behaves. Good instructions are the single biggest lever in shaping a custom agent.
The editor
The Instructions section is a plain text editor. Markdown formatting in the prompt is fine — the model reads it, even though the editor doesn't render it. Instructions auto-save to draft on every edit.
For catalog agents, edits add to — rather than replace — the catalog's base instructions.
What to put in instructions
- Role — what the agent is and what it's responsible for.
- Audience — who it talks to and how formal or informal it should be.
- Constraints — what it must not do. For example: "never auto-send email", "never modify CRM records without confirming first".
- Tone and style — formality, preferred response length, structure preferences.
- Tool guidance — when to reach for which tool. Especially helpful for
search_kbandsearch_org_contacts. - Examples — one or two short examples of the kind of response you want.
Refine with AI
Refine with AI opens a dialog that rewrites your instructions for clarity and structure. Provide your current instructions and an optional refinement prompt — for example, "make it more concise" or "add safety constraints". The dialog shows the proposed rewrite. Nothing is overwritten until you accept it.
How instructions interact with tools
Instructions and tool descriptions both influence what the model does. Tools have their own descriptions; the model uses those to decide when to call them. If an agent isn't calling a tool you expect, add explicit guidance in the instructions pointing it toward that tool.