How to write a prompt that actually works.

Many disappointing AI answers are not the tool's fault. They are the result of an ask that no colleague could have answered well either. "Write me a marketing email" gets you something resembling the average of every marketing email you have ever received, because you gave it no reason to produce anything more specific.

There is no magic wording and no secret phrase. A prompt that works is just a good brief, the kind you would give a capable new starter on their first day. Here is the shape of one.

Briefing well is the first step of a longer loop, not the whole method. What happens after the answer comes back, challenging it and deciding what you will put your name to, matters just as much, and we set that out in closing the loop. This note is about the step you start with.

The four parts of a good brief

• The task. What you want, specifically, including any boundaries. "Draft a reply declining this request but keeping the door open for next quarter, and do not offer a discount" beats "reply to this email".
• The context. What it needs to know: the email thread, the previous proposal, who the reader is, the price you actually charge. Attach it, paste it, or point the tool at it.
• The shape. What the output should look like: length, format, structure. "Three short paragraphs, no bullet points, sign off as Sarah" saves a round of edits.
• The tone. Formal or friendly, cautious or direct. Better still, paste an example of something you have written that sounds right, and say "like this".

You will not need all four every time. This is a checklist, not a formula. Use the parts that remove uncertainty; a simple factual question does not need a miniature specification.

Notice what is missing. There is no "act as a world-class copywriter". The useful words are the facts, the constraints and the result you want.

Two worked examples

First, writing. The vague version:

"Write a follow-up email to a customer."

The version that works:

"Draft a follow-up email to a customer who received our quote two weeks ago and went quiet. We quoted £4,800 for a kitchen fit starting in March. I want to nudge without pressuring: check the quote landed, offer to answer questions, and mention March is filling up, which is true. Do not invent a discount, deadline or other incentive. Friendly but professional, three short paragraphs, sign off as Finlay. Here is the original quote email for reference: [paste]"

Same tool, thirty seconds more effort, and the second one comes back nearly ready to send. Notice that most of the extra length is not clever wording. It is information.

Now analysis, because none of this is only for writing. The vague version:

"Analyse this spreadsheet."

The version that works:

"Review the attached monthly sales spreadsheet for the leadership team. Compare the latest quarter with the previous one, identify the three changes most likely to matter commercially, and show the figures behind each conclusion. Return a short summary followed by a comparison table. Distinguish facts from your interpretation, and flag any missing data rather than filling the gaps yourself."

Same four parts, different job. The last sentence is the one that earns its keep. Told to flag the gaps, it tells you what it could not find. Left to itself, it fills them.

Then talk to it

The first answer is a first draft, not a verdict. The real skill is the follow-up: "shorter", "less formal", "you have promised something we did not agree, take it out", "give me three versions of the opening line". People who treat AI as a conversation get far more from it than people who treat it as a slot machine, re-rolling the same prompt and hoping.

Two habits that pay for themselves

• Save what works. When a prompt produces something good, keep it. A shared note of five proven prompts for your team's routine jobs beats any prompt library off the internet, because it is about your work.
• Stop repeating yourself. If you keep typing the same background, about your business, your tone, your prices, move it somewhere permanent: the tool's standing instructions, or a project where the documents live once. Context you set once is context nobody forgets to paste. Keep it current, though: an old price list supplied automatically is worse than a new one you occasionally forget to attach. We wrote more about that in why the same AI works better for one team than another.

The short version

Brief it like a person. Give it the task, the context, the shape and the tone. Then edit by talking, not by starting again. That is most of what "prompt engineering" amounts to. A team can understand it in an afternoon; turning it into a habit takes practice. It is where our training starts.

Written with Claude, then steered, cut and checked by me. I answer for it.

Finlay Carmichael

Finlay Carmichael is the founder of The AI Lab, helping teams put AI to practical use through hands-on training and clear advice. He brings nearly 30 years' experience building and leading a Microsoft Business Solutions partner, working with hundreds of clients across the UK. Request a call