How to Write AI Prompts That Actually Work

Most of the time, when someone tells me “I tried AI and it wasn’t useful,” the model wasn’t the problem. The prompt was.

A language model does exactly what you ask — the trouble is that most of us ask the way we’d ask a colleague who already knows the background. The model doesn’t. It only knows what you put in front of it. Once you treat a prompt as a short, specific brief instead of a quick question, the quality jumps immediately. No new tool required.

Here’s the structure, the mistakes that waste the most time, and real before-and-after examples you can use today.

A prompt is a brief, not a question

Think of the model as a fast, capable freelancer on their first day. They’re skilled, but they know nothing about your business, your customer, or what “good” looks like to you. A useful brief covers up to five things:

1. Role — who the model should act as (You are an experienced UK small-business accountant…).
2. Context — the situation and any facts it needs (I run a six-room guesthouse in Cornwall; our busy season is…).
3. Task — exactly what you want done. One clear job.
4. Format — how the answer should be structured (Reply as five bullets, each under 20 words).
5. Constraints — the rules and limits (British English. No jargon. Don't invent numbers.).

You won’t always need all five, but the more you include, the less the model has to guess — and guessing is where bad answers come from.

The anatomy of a good prompt — a brief, not a question.

Before and after

Here’s a weak prompt, the kind most people write:

Write a marketing email for my café.

You’ll get something — generic, over-excited, full of “Indulge in…” and “Nestled in the heart of…”. Now the same request written as a brief:

You are a friendly copywriter for a small independent café.

Context: We're a café in Bristol. This week we launched a new
weekend brunch menu (served 9am–1pm, Sat & Sun). Our customers
are local, 25–45, and dislike anything that sounds corporate.

Task: Write a short email announcing the brunch menu to our list.

Format: A subject line + 120–150 words. One clear call to action:
book a table.

Constraints: Plain British English, warm but not gushing. No emojis.
Don't invent menu items or prices — leave [brackets] where I fill in.

The second one gives you something you can actually send after a small edit — the right length, the right tone, and honest placeholders instead of invented facts.

The mistakes that waste the most time

• Too vague. “Make it better” gives the model nothing to aim at. Say what better means: shorter, warmer, more specific, fewer claims.
• No context. The model can’t read your website or your inbox. If a fact matters, paste it in.
• Asking for everything at once. “Write my whole marketing plan” gets you a shallow answer. Break big jobs into steps and build them up.
• No format. If you don’t say how you want the answer, you get an essay. Ask for a table, a list, a word count.
• Accepting the first draft. The first answer is a starting point, not the finish. The real gains come from the next three messages.

Show it what “good” looks like

One of the most powerful moves is to give the model an example. If you have a piece of writing you like — a past email, a competitor’s product page, your own tone of voice — paste it in and say “match this.” This is called few-shot prompting, and it usually beats describing a style in words:

Here are two product descriptions we've written before that we like:
[paste examples].

Write a new one for this product in the same voice and length:
[product details].

Treat it as a conversation

You don’t need the perfect prompt first time. Some of the most useful replies come from following up:

• “That’s too formal — rewrite it the way you’d text a friend.”
• “Give me three versions with different opening lines.”
• “What’s weak about this draft? Fix the weakest part.”
• “You invented a statistic there. Remove anything you can’t support.”

That last one matters. Models will confidently make things up. Tell it not to invent figures, ask where a claim comes from, and check anything factual yourself.

A worked example: a rough note into a client email

Say you’ve come off a call and jotted this down:

call w/ Sarah — wants garden landscaping done before her daughter's
wedding, 12 Sept. budget tight-ish. worried about patio drainage.
send quote + reassure re: timing

Here’s the brief:

You are writing on behalf of a small landscaping firm. Turn my rough
notes into a warm, professional follow-up email to a client, Sarah.

Notes: [paste the notes above]

Task: Confirm we can complete before 12 September, acknowledge the
patio-drainage concern and say we'll assess it on the site visit, and
note that a detailed quote will follow separately.

Format: Under 150 words. Friendly and reassuring, plain British English.
Leave [brackets] for anything I must fill in (like a site-visit date).
Don't promise a price or a firm date I haven't given you.

You get a sendable draft in seconds — one that reassures Sarah on the two things she actually cares about (timing and drainage) without inventing a price or a commitment you never made. Two minutes, instead of fifteen staring at a blank screen.

When a better prompt will not fix the problem

Prompting matters, but it is easy to blame the prompt for a task the model cannot reliably complete. Stop rewriting the instruction and change the workflow when:

• The required fact is missing. A model cannot infer your real price, deadline or policy. Supply an approved source or leave a placeholder.
• The source is weak. A beautifully structured prompt applied to outdated notes still produces an outdated answer.
• The task needs a current lookup. Ask for sources and verify them, or use a system designed to retrieve the relevant documents.
• The decision is high-impact. Legal, medical, financial and employment decisions need qualified human judgement, not merely a more detailed prompt.
• The output cannot be tested. If you cannot describe what “correct” means, neither a person nor an AI system has a dependable finish line.

This distinction becomes more important as AI tools move from producing text to taking actions. A writing prompt can produce a poor draft; a connected agent can send it, update a record or trigger another system. The UK Government AI Playbook recommends meaningful human control at the stages where impact and risk justify it. Good prompting helps, but permissions, review and rollback are separate design decisions.

A five-minute prompt quality check

Before you run an important prompt, check it against this list:

1. One task: is there a single main action, rather than five competing requests?
2. Known facts: have you separated supplied facts from things the model would have to guess?
3. Audience: does the model know who will use or read the result?
4. Finish line: have you specified length, structure and what a good answer must contain?
5. Boundaries: have you said what it must not invent, reveal, decide or do?
6. Verification: do you know which claims, calculations or decisions a person will check?

If the first answer is weak, diagnose the failure. “Too long” needs a format constraint. “Generic” needs examples or specific evidence. “Wrong” needs a better source or a different tool — not more adjectives.

A reusable template

Keep this somewhere and adapt it each time:

Role: You are a [role / expertise].
Context: [who I am, the situation, key facts].
Task: [the one thing I want done].
Format: [structure, length, style].
Constraints: [language, tone, do's and don'ts, "don't invent facts"].

Five starters for common jobs

Adapt these and keep them handy:

• Summarise: Summarise the text below for a busy [role]. Give me the 5 key points as bullets, then one line on what to do next. Text: [paste].
• Improve writing: Rewrite this to be clearer and warmer — same meaning, 20% shorter, plain British English. Keep any facts exactly as written: [paste].
• Brainstorm: Act as a sceptical [role]. Give me 10 ideas for [goal], then mark the three you'd actually try and say why.
• Analyse: Here's some data: [paste]. In plain English, what are the three most useful things it tells me — and what would you double-check before trusting it?
• Learn something: Explain [topic] to me like I run a small business, not a computer-science class. Use one real example and tell me the one thing most people get wrong.

The one skill worth practising

Good prompting is the single habit that makes every other AI tool more useful — for writing, research, planning, analysis and coding. It comes down to three things: be specific, give context, and iterate instead of accepting the first draft.

If you’d like to practise this on your own real work — your emails, your content, your data — that’s exactly what we do one-to-one in a private ChatGPT training session, or you can book a private lecture and bring whatever you’re stuck on.