Prompt engineering

What is prompt engineering?

Writing a prompt that works once is easy. Getting it to work for a hundred different customers, badly worded questions, people in a hurry and odd edge cases is another matter. That’s prompt engineering: treating a model’s instructions like a piece of software that you design, test and maintain.

It isn’t magic or a bag of tricks. It’s closer to writing a good internal procedure: what needs doing, with what information, in what format, and what’s off limits. Except that this “employee” takes everything literally.

How it works

The work usually follows a cycle:

1. Define the goal. What a good answer looks like and what a bad one looks like, with examples.
2. Write a first version. Often in the system prompt: role, tone, rules and limits.
3. Test with real cases. Common questions, unusual ones, angry customers, attempts to pull the model off task.
4. Adjust and repeat. Change one thing at a time and compare.

The most common techniques are giving examples (few-shot), asking the model to reason step by step before answering (chain of thought), fixing an output format and keeping instructions clearly apart from data. For answers about your own business, it’s combined with RAG, which passes your documents to the model.

Example

An electrical wholesaler in Bristol uses AI to sort incoming email into order, issue, quote or invoice. The first version of the prompt keeps mixing up issues and urgent orders. After going through twenty misfiled emails, the team settles on this:

Classify the email into ONE category: order, issue, quote or invoice.
Issue = something already delivered that is faulty or missing.
If the customer asks for new stock, it's an order, even if they say "urgent".
If you're unsure between two, reply "review".
Return only the category, in lower case.

The rules come straight from real mistakes. And the “review” option stops the model forcing an answer when it isn’t sure.

What it means for your business

You don’t need a prompt engineer to use an AI agent. What you do need is to review its work early on and correct it with specific examples: “this call should have been transferred”, “we don’t quote that price over the phone”.

UTIO is built around that cycle. Creator builds the agent by reading your website and it starts in draft mode, so you check things before anything is sent. If we set it up for you, we tune the script with real conversations during the first month. For your own projects, the AI API lets you try out prompts to generate and classify text.