Prompt

What is a prompt?

Every time you type into an AI assistant, you’re writing a prompt. It can be a short question (“what’s an IVR?”), a command (“summarise this email in three lines”) or a brief with material attached (“using these notes, draft a quote”).

The model can’t read your mind. All it sees is the text it receives, broken into tokens. That’s why two prompts that look the same to you can produce very different results. The clearer the request, the less the model fills in on its own.

In a real application, what reaches the model isn’t just what the user types. It’s combined with a system prompt holding the fixed rules, the conversation history and sometimes extracts from documents. All of it has to fit inside the context window.

How it works

A good prompt usually has four parts:

1. Task. Exactly what you want: answer, summarise, classify, write.
2. Context. Who the answer is for, what information to use, what the situation is.
3. Format. Length, tone, structure: a list, a table, one paragraph.
4. Limits. What it mustn’t do: make up prices, promise delivery dates, wander off topic.

If you also show it one or two examples of the answer you expect, you’re using few-shot prompting. And the craft of testing and refining prompts until they work reliably is called prompt engineering.

Example

A physiotherapy clinic in Leeds wants AI to help it reply to Google reviews. The prompt “reply to this review” gives bland, generic text. This one works much better:

You are the front desk of a physiotherapy clinic in Leeds.
Reply to this Google review in no more than 3 sentences.
Warm, friendly tone, British English, no emojis.
If the review is negative, apologise and invite them to ring the clinic.
Don't offer discounts and never discuss diagnoses.
Review: "{review text}"

It has a task, context, a format and limits. The reply comes out shorter, more consistent and promises nothing the clinic doesn’t want to deliver.

Prompts at UTIO

With UTIO you don’t need to write prompts to get the agent working. Creator builds the agent by reading your website, in one minute, and the agent starts in draft mode: it prepares the reply and a person checks it before it goes out. You see how it answers before you give it more autonomy.

If you’d rather we set it up, we tune its script with real conversations during the first month. And if you’re a developer, the AI API lets you send your own prompts to generate and classify text.

Knowing what a prompt is makes it easier to review what an AI agent does and to ask for specific changes.