Skip to main content
The system prompt is the agent’s brief. Every reply the agent writes starts from it. Settings shape how the agent talks; the prompt decides what it says and when it acts.

​
What a good chat prompt covers

Who the agent is, which business it speaks for, and what it is allowed to talk about. Name the boundary explicitly — “you only handle questions about orders and delivery; anything else, say you will pass it to the team.”
Messaging is not a phone call. Say how long replies should be (“two or three short sentences”), how formal, and which language to reply in — usually the language the customer wrote in.
The single most common reason a function is never used is a prompt that never mentions it. For each function, say the situation that should trigger it: “when the customer asks for the price list, call send_price_list.”
Negative instructions are as important as positive ones. No competitor comparisons, no discounts you have not authorised, no medical or legal advice, no promises about delivery dates.
Describe the states that count as done — order placed, question answered, customer said they are not interested. The follow-up stop condition can refer back to this, so the agent stops chasing conversations that are actually finished.

​
Example skeleton

​
Prompt tips specific to chat

Write the trigger, not the capability. “You can send documents” tells the model nothing about when. “When the customer asks how to claim warranty, call send_warranty_pdf” does.
Remember the agent only sees the last N messages. Context Length defaults to 10. If something must persist for the whole relationship — a customer’s plan, their language preference — read it from the CRM contact in a function rather than relying on it staying in the window.
Say what to do when it does not know. Without an instruction, a model fills the gap. “If you are not sure, say you will check and call schedule_callback” is better than a confident invention.
The prompt is not a security boundary. Anything the agent can technically do — every function you attach — it may do if a conversation talks it into that. Do not attach a function whose worst-case use you would not accept.

How to write a good prompt

The voice-agent prompt guide — most of it applies here too

Functions

What you are describing triggers for