Skip to main content

AI and Automation

Prompt Engineering

How you ask genuinely changes what you get. It is a skill worth having in-house rather than a profession worth hiring, and it should never be the whole product.

Definition

Prompt engineering is writing the instruction given to a model so that it produces what you want.

What It Actually Means

The effect is real and larger than people expect. The same request phrased two ways can produce a usable answer and a useless one, because the model is responding to everything in the input, including what you did not realise you were signalling.

What Genuinely Works

The techniques that hold up are unglamorous and mostly amount to communicating clearly.

Be specific about the output. Length, format, audience, what to leave out. Models default to a generic middle when not told, and most disappointment is an unstated expectation.

Show an example. One or two examples of a good answer does more than several paragraphs of description. This is the highest-return technique by a distance.

Give the role and the context. Who is asking, what for, what they already know. It shifts register and level appropriately.

Say what not to do. Constraints are followed reasonably well when stated and not at all when assumed.

Ask for reasoning on hard tasks. Requiring the model to work through steps before answering measurably improves anything involving logic, because the intermediate steps become part of what it is predicting from.

Break big requests up. Several focused steps beat one instruction carrying four jobs.

Where It Is Oversold

Two things are worth being sceptical about.

Prompt engineering as a career. It was briefly presented as a profession with a salary attached. In practice it is a skill that anyone using these tools picks up in a few weeks, and it is better distributed across the people doing the work than concentrated in a specialist. Newer models also need less coaxing, which is steadily reducing the specialist premium.

Prompt engineering as a product. If a supplier’s offering is mostly a well-written prompt in a wrapper, you are paying for something you could write, and it will not hold up. Prompts are trivially easy to change and impossible to defend commercially.

The durable version is a system: what information is retrieved and supplied, what the output is checked against, what happens on failure, where a person sits. The prompt is one component inside that, and it is the least valuable one.

What To Ask

  • Are the prompts version-controlled and tested? They are configuration that changes behaviour. Editing them in production, untested, is the norm and it should not be.
  • How do we know a change improved things? You need a set of real examples with expected outputs, run before and after. Without it, prompt changes are guesswork that feels like progress.
  • What happens when the model updates? Prompts tuned for one model can behave differently on the next. Tests are how you find out before your customers do.
  • What is the product if the prompt is removed? If the answer is nothing, reconsider the price.

The Honest Summary

It is worth learning, and there is nothing mysterious about it. An afternoon of deliberate practice will get most people most of the benefit, and the remaining gains come from the engineering around the prompt rather than the wording of it.

Where prompting stops being enough is when the model needs to know things. No amount of careful phrasing gives it access to your data, which is a retrieval problem instead.

More terms are in the glossary.

Portrait of Alexander De Sousa, founder of Digital Royalty
Founder-led
“I’ve put everything I know into how this company works — the standards, the method, the care on every project. It runs through the whole team, and I hold us all to it.”

Alexander De Sousa · Founder LinkedIn

Featured on BBC Radio Solent

Get started

Tell us what you need

A few quick questions, then a straight answer from a real person — usually within a few hours.

What happens next

  1. 1 We reply A real person reads your answers and gets back to you, usually within a few hours.
  2. 2 Questions or a short call If anything isn't clear, we ask by email or book a short call. Before any call we audit your website and make the quick fixes it finds, free.
  3. 3 Your fixed quote A clear, fixed price in writing, within 48 hours of any call. There is no obligation to go ahead.

Tell us what you're working on

Whether it's a new site, a platform, or a process that shouldn't be manual any more — we'll tell you honestly if we can help.