Who should read this article: Administrators
The Prompt is where you tell an AI agent who it is, what it should do, and what it must never do. This article explains how to write one, and how to improve it once the agent starts taking calls.
Introduction
An AI agent gets its behavior from two places: the Prompt you write, and the knowledge bases you assign to it. Keep the two apart and both stay easy to maintain:
- The prompt holds behavior — role, tone, rules, and when to hand over to a human.
- A knowledge base holds facts — policies, prices, product details. Refer to Knowledge bases for more information.
Put a price in the prompt and you have to edit every agent when it changes. Put it in a knowledge base and you update one document.
A prompt is never right the first time. Write a short version, test it in the Test panel, listen to what the agent actually says, and fix what went wrong. Refer to AI Agents for more information about the Test panel.
Starting from a template
The Prompt template field offers three ready-made prompts:
| Template | Use it for |
|---|---|
| Receptionist | Greeting callers, understanding what they need, and answering general questions from the knowledge base. |
| Support | Helping users understand and resolve issues, with a separate mode for working through a problem. |
| Outreach | Placing outbound calls to inform users, with a defined call purpose and a closing sequence. |
| Custom | Writing the prompt yourself, from an empty field. |
Selecting a template fills the Prompt field. Edit that text to suit your business — the template is a starting point, not a finished prompt.
Start from the template closest to your use case even when the fit is rough. The templates carry the structure and the safety rules that take longest to get right, and editing wording is quicker than building both from nothing.
What a good prompt contains
The supplied templates are organized into labelled sections, and a prompt you write yourself benefits from the same shape. Give the agent these, in roughly this order:
| Part | What to write |
|---|---|
| Role | Who the agent is and what it is the first point of contact for. One or two lines. |
| Job | The two or three things it is there to do. |
| Rule priority | Which instruction wins when two conflict — safety rules above accuracy, accuracy above style. |
| Operating modes | The distinct jobs it switches between, such as answering questions and working through a problem, and what it does in each. |
| Information reuse | Reuse what the contact already said, and never ask for the same detail twice. |
| Communication style | Tone, sentence length, and how to respond to an upset contact. |
| Handoff conditions | The situations where it stops and transfers to a human. |
| Strict rules | The short list of things it must never do. |
Be specific about limits
The strict rules are the part that keeps an agent out of trouble, and the part most often left too vague. Write them as flat prohibitions:
## STRICT RULES
* Never invent information, policies, terms, or conditions
* Never confirm account-specific data or transaction status
* Never use information outside the knowledge base
* Do not promise actions or outcomes you cannot deliver
Say when to hand over
An agent that never transfers is worse than no agent. Name the situations plainly — an upset contact, a request for a human, anything needing account access the agent does not have — and set the destination in the Handoff section of the agent.
Keep it short and structured
Short, structured prompts work better than long prose. Headings and bullets beat paragraphs. Say the same rule once. The Prompt field holds up to 20,000 characters, but a prompt that long is usually a sign that facts have crept in that belong in a knowledge base.
Writing for voice
An AI agent is heard, not read. That changes what a good instruction looks like:
- Ask for short answers. "Keep responses to two or three sentences" prevents monologues.
- Tell it to ask one question at a time. Two questions in one breath get one answer.
- Tell it to confirm what it heard before acting on a number, a date, or a spelling.
- Set the language explicitly if your contacts expect one, including whether the agent may switch.
- Avoid anything that only works in writing — lists to read out, links, formatting.
Testing and improving
- Write the shortest prompt that describes the job.
- Open the Test panel and click Start conversation.
- Ask the questions your customers actually ask, and a few the agent should refuse.
- Note where it guessed, rambled, or missed a handoff, and fix that one thing.
- Test again.
Change one thing at a time. Changing several at once tells you the agent improved but not which edit did it.
Each conversation starts fresh. The agent has no memory of earlier calls and reads the whole prompt and knowledge base from the beginning every time, so anything it needs to know has to be in one of those two places.
- Re-test after any large prompt edit, not only after the first draft.
- Rewriting a prompt with an AI writing tool is fine, but read the result before you save it — rewrites quietly soften business rules and compliance wording.
- Keep a copy of a prompt that works before you experiment on it.