Skip to content

Prompt ​

The prompt is the text that defines who answers and how the conversation is led. It is the part that changes the agent's result the most — more than any other setting on the screen.

By the end of this guide you will know how to write a prompt that sounds like your company, without repeating information that already lives in other topics.

Before you start ​

Be clear about: who the person answering is (name and role), what the company does in one sentence, and how it talks to customers — formal or informal, short or long messages.

What to write ​

Write it the way you would explain it to someone who joined the team today. Three blocks are usually enough:

Who it is. Name, company, city and what the company does.

You are Linda, a salesperson at Lindóia Alianças, in Porto Alegre. The store buys gold from people who want to sell, with a free in-person appraisal, and sells gold jewelry and original Ray-Ban glasses.

How it opens the conversation. The first message is the one that goes wrong the most when it isn't written down.

Your first message has two parts and nothing else: you introduce yourself by name and say where you're speaking from, and you ask an open question. If the person already arrived saying what they want, don't repeat the question — acknowledge what they said and move on.

How it leads. Tone, pace, message length.

You lead the conversation; you don't wait. Talk like a store person: warm, curious, short messages, one question at a time. Brazilian Portuguese, at most one light emoji.

Use headings to separate the instructions ​

Start each block with # and a heading. This isn't decoration: the agent reads each section as a topic of its own, and the instruction from one block stops bleeding into the other. Without headings, the text becomes a single block and the rules get mixed up.

# Who you are
...

# How you open the conversation
...

# How you lead
...

The bar above the field has the formatting buttons, and the Sections in the text column, on the right, shows the structure you created — indented by level. If it is empty, your text has no sections.

There is a second, less obvious reason: when the text goes past the recommended size, it is cut from the end. In a sectioned text you lose the whole last section, not half a sentence in the middle of a rule.

What NOT to write here ​

The right-hand column reminds you of this while you write:

  • Address, hours, price and links — go in Knowledge → Fixed facts. Written in the prompt, they are harder to maintain and compete for space.
  • When to transfer to a person — goes in Transfers.
  • What it must never do — goes in Restrictions, just below the prompt, which take priority over this text.

Using contact data in the text ​

Type {{ anywhere in the text — or click the { } button on the bar — and pick a variable. It is replaced by the real value at conversation time.

You are talking to {{first_name}}, from the company {{contact.company}}.

The picker shows three groups:

  • Contact — name, email, phone, company, job title, city.
  • Custom fields — the fields your account created.
  • From this webhook — only shows up when the agent has the webhook trigger on, and lists the fields from your own payload, with the test value next to them.

A variable that doesn't resolve stays blank

If the data doesn't exist in that conversation, the variable becomes empty text — the sentence breaks silently. Use the {{field|default text}} format to set what shows up when the data isn't there: {{first_name|you}}.

How to know it worked ​

  • The Sections in the text column lists the headings you wrote.
  • The Length meter, at the top of the panel, is far from the limit.
  • In the menu, Prompt shows a ✓ and the summary shows the section count.
  • With the Test button, the agent's first message sounds the way you wrote it.

Common problems ​

The agent ignores part of what I wrote. Cause: the text went past the recommended size and the excess is cut at answer time, always from the end. How to check: the Length meter turns amber and a warning appears at the top of the panel. Solution: trim it down, or move fixed data to Knowledge — that's what it's there for.

The variable shows up empty in the conversation. Cause: the contact doesn't have that data, or the variable is a webhook one in a conversation that didn't start from a webhook. How to check: an amber banner appears in the panel listing the webhook variables at risk. Solution: use the format with a default, {{field|default text}}.

The agent introduces itself again in the middle of the conversation. Cause: the prompt describes the opening without saying it applies only to the first message. Solution: make it explicit — "Your first message has two parts" — and add what to do when the person has already said what they want.

GetRaze - AI-Powered Lead Generation