Prompt engineering basics: how to write prompts that get better results

You already use AI. It drafts your emails and summarises your reading, and the output comes back fine. Fine is the problem: generic, hedged, close to what you asked for but never quite what you needed.

Usually the model isn't the limit. Your prompt is.

Prompt engineering is the skill of asking properly. It needs no code and no training data, and you can get better at it this afternoon. Change the question, and the answer changes with it.

Why the prompt does the heavy lifting

Most people treat a prompt like a search query. A few keywords, then disappointment when the result misses.

A prompt is a brief, not a query. The same model, on the same day, will hand you something useless or something you can send, depending on how much of that brief you bothered to write.

Think about handing the task to a new colleague. You wouldn't say "write something." You'd tell them who it's for, roughly how long, what tone, what to leave out. AI needs the same input for the same reason. It has huge general capability and no knowledge of your situation, and everything about your situation has to come from you.

Seven techniques worth learning

1. Be specific enough to feel silly

Vague in, vague out. This is the one change that fixes most bad outputs.

The second version names the platform, the length, the tone, the one fact that matters and the thing to avoid. You get something usable first time instead of three rounds of back and forth.

The same move works anywhere:

• "Summarise this article" becomes "give me the three main findings as bullets, in language a non-expert would follow".
• "Help with my cover letter" becomes "rewrite this paragraph to sound more confident and less apologetic, under 100 words".
• "Analyse this data" becomes "compare these Q1 and Q2 figures, flag anything unusual, and write it for a finance director".

2. Give it a role

Tell the model who it is before you tell it what to do. That single line shapes vocabulary, depth and what it decides to mention.

A role narrows the focus. A financial analyst digs into the numbers. A technical writer strips the jargon. A recruiter skims for red flags. Same question, different lens, and you get to pick the lens.

3. Show one example

Describing the style you want is harder than showing it. Paste in a sample of the output you like, then ask for more of the same.

This is few-shot prompting, and one example does most of the work. It doesn't need to be polished. Real copy from your own site beats any description of your brand voice, because the model picks up sentence length, word choice and where you put the emphasis without being told.

4. Name the format you want

If the output has to slot into something, say what shape it should be.

Ask for a table, bullets, JSON, a two-column layout, whatever you need. It costs you one sentence and saves you the reformatting.

5. Ask it to work through the problem first

For anything involving logic, comparison or numbers, add a line telling the model to reason before it concludes: "work through this step by step, then give your answer."

It sounds like a trick. It isn't. Forcing the reasoning into the open makes the model less likely to leap to a confident wrong answer, and you get to see where it went off course instead of guessing. Worth doing when you're checking whether a set of figures supports a claim, or picking between options with real trade-offs. Skip it for simple rewriting, where it just pads the response.

6. Let it ask you questions, and let it say "I don't know"

Two lines you can bolt onto almost any prompt.

The first: "Before you answer, ask me up to three questions about anything you need to know." Instead of guessing at your context, the model tells you what's missing from your brief. It's the quickest way to learn what a good prompt for that task looks like, because the questions are the gaps.

The second: "If you're not sure about something, say so rather than guessing." Models default to sounding confident whether or not they should. Given explicit permission to hedge, they'll flag the shaky parts, and you know where to check. Worth adding to anything involving figures, dates or names.

While you're there, paste in the source material. A model working from your actual document is far more reliable than one working from memory. Attach the report, the transcript, the competitor's pricing page, then ask your question about it. Most of what looks like a model problem is a missing-context problem.

7. Treat the first answer as a draft

Your first prompt won't be right. Nobody's is. The fix is to say what's wrong rather than starting again from scratch.

• "Close. The tone is too formal. Redo it the way you'd explain it to a colleague."
• "Good structure. Cut the last paragraph and add a concrete example to point two."
• "You've assumed we sell to consumers. We sell to hospitals. Try again."

Each round is cheap. Two rounds of correction almost always beat one heroic prompt.

A skeleton you can reuse

When a prompt matters, fill in five slots.

• Role: who the model should be.
• Task: the one job, stated as a verb.
• Context: what it can't know about your situation.
• Format: the shape of the output.
• Constraints: length, tone, anything off limits.

Put together:

Write a few of these and you'll start doing it in your head.

Where prompts usually go wrong

Telling it what not to do. "Don't be formal" gives the model something to avoid and no target. "Write it the way you'd text a colleague" gives it something to aim at. Positive instructions land better.

Assuming it knows your context. It doesn't know your audience, your pricing, your brand voice or last week's conversation unless you put it in the prompt. If your copy is irreverent, say so. If you're writing for compliance officers, say that too.

Burying the instruction. Politeness is fine, hedging isn't. "I was wondering if you could maybe pull out a few thoughts" leaves the model guessing at what you want and how much of it. "Give me five bullet points" doesn't.

Stacking three jobs into one prompt. Research, then draft, then critique gets you something mediocre at all three. Split them into separate prompts, or number the steps explicitly.

And when a result comes back wrong, resist switching tools. Adjust the prompt first. Far more often than not, the question was the problem.

When better prompts are enough, and when they aren't

Fine-tuning gets a lot of attention. Most people don't need it, and almost nobody needs it yet.

Stick with prompting when:

• The task changes a little each time.
• You're working with a general model like ChatGPT or Claude.
• You want to test an idea before spending money on it.
• You're at roughly 70% and need the last stretch.

Look at fine-tuning when:

• You run the same narrow task hundreds of times.
• The model gets it wrong the same way every time, however you ask.
• You have domain data worth training on.
• You've genuinely exhausted prompting and can measure what's still missing.

Better prompts are free and immediate. Fine-tuning costs time, data and money. Start with the cheap thing and see how far it gets you. When you have hit the ceiling, our step-by-step guide to fine-tuning an LLM for your specific use case picks up exactly where prompting leaves off.

Try it on something real

Pick one task you do most weeks. Write the prompt the way you normally would, then write it again with a role, the context you'd forgotten, and the format you want back. Run both.

The difference won't be subtle. Then do it again tomorrow, on something else. That's the whole skill. Not cleverness, just the habit of saying what you want.

Frequently asked questions