Writing a system prompt

What the system prompt is for

The system prompt sets tone and behavior: who the agent is, how it should address customers, what it should never do (e.g. quote prices it isn't sure of, or promise a refund). It is not where facts, policies, or product details should live — put those in Knowledge instead, where they can be retrieved, cited, and kept up to date without touching the prompt.

A good rule of thumb: if a fact would go stale (a price, a policy, a turnaround time), it belongs in Knowledge. If it's a standing instruction about how the agent should behave, it belongs in the system prompt.

Rewrite with AI

If you have a rough one- or two-line description of the agent's role, the Rewrite with AI button in Settings expands it into a fuller prompt. You can Undo back to what you had before accepting the rewrite. Treat the output as a draft — read it before saving and adjust anything that's too generic for your business.

Example

A reasonable starting prompt:

You are a support agent for Acme Outdoor Gear. Be warm, concise, and honest — if you don't know something, say so and offer to connect the customer with a human rather than guessing. Never quote a price or shipping date that isn't in your knowledge base.

Notice what's not in there: no actual prices, shipping policy text, or return windows — those live in Knowledge, where the agent retrieves and cites them per question, and where you can update them without redeploying the agent's behavior.

Where behavior settings live instead of the prompt

Some behavior is controlled by dedicated settings rather than prose in the prompt, because they change how the platform executes rather than what the model says:

• Whether the agent will only answer with supporting evidence, or can fall back to general knowledge — see Retrieval & grounded answers.
• Whether the agent can escalate to a human or mark a conversation resolved — see Escalation to a human.