In this article
- What is AI prompt engineering?
- 1 Give it a job title
- 2 Say what it is for, and who is reading
- 3 Paste the real material
- 4 Ask for the shape of the answer
- 5 Show it one example
- 6 Set the boundaries
- 7 Make it interview you first
- 8 Split big jobs into steps
- 9 Make it check its own work
- 10 Keep a prompt library
- Which tip should I use first?
- What mistakes do people make with prompts?
- Prompt engineering: questions people ask
The short answer
AI prompt engineering is the skill of writing instructions that get a useful answer first time. The ten tips that matter most: give the AI a role, state the purpose, paste real material, choose the format, show an example, set limits, let it ask questions, split big jobs, make it check itself, and save what works.
Most people use AI like a vending machine. Press a button, get a snack, complain about the snack. Prompt engineering is the dull-sounding name for getting better snacks, and it comes down to what you type before the AI starts typing back.
These are the ten habits I teach first, in roughly the order they pay off. Every one comes with a prompt you can copy into ChatGPT, Claude, Copilot or Gemini. Mix them. Three of these in one prompt will beat any single tip done perfectly.
What is AI prompt engineering?
AI prompt engineering is the practice of writing instructions for an AI model so it gives you something usable on the first or second attempt. It is closer to briefing a capable new starter than to coding. You say who it is, what you want, what it should work from, and what good looks like.
There is no secret syntax. The people who get the best results are the ones who explain themselves clearly, which is why a good manager often beats a good programmer at this. If you want the longer version, I cover it in what is prompt engineering for business and in the full guide to how to write prompts for AI.
Give it a job title
An AI with no role gives you the average of everything it has read. An AI with a role gives you something that sounds like it has done the job. Name the profession, the experience level and, if it helps, the attitude.
Write a client email about a late payment.
You are a credit controller at a UK accountancy firm. You have chased invoices for twenty years and never lost a client doing it. Write an email to a client whose invoice is 30 days overdue. Firm, polite, and it should leave the relationship intact.
Same task. The second one comes back sounding like someone who has had this conversation before.
Say what it is for, and who is reading
A summary for your board and a summary for your team are different documents. The AI cannot know which one you need unless you say. Give it the audience, what they already know, and the one thing they have to decide.
Summarise the report below for our board. They have five minutes and one decision to make: whether to renew the supplier contract. Lead with your recommendation, then give the two strongest reasons. Assume they have not read the report. [paste report]
Paste the real material
If you ask for a post about your new service, you get a post about any service. Paste your notes, your old emails, the customer's actual message. The AI is excellent at reshaping what you give it and mediocre at guessing what you meant.
Below are my rough notes from this morning's client call. Turn them into a recap email of four bullet points I can send today. Use only what is in the notes. If something looks like a decision but is not clearly stated, put it under 'To confirm'. [paste notes]
Why it works: the last sentence stops the AI inventing agreements that never happened.
This is also the safest habit. Paste the material you are allowed to paste, and read how to keep information safe using AI tools before you drop in anything with client names in it.
Ask for the shape of the answer
Tables, bullets, a three-line email, a script with timings. If you do not choose the shape, the AI chooses a wall of friendly paragraphs, and you spend ten minutes extracting the useful bit.
Compare the three suppliers in the attached quotes. Put the result in a table with these columns: supplier, monthly cost, contract length, biggest risk. Below the table, tell me which one you would pick in a single sentence.
Asking for the format first also makes errors easier to spot. A blank cell in a table is much harder to hide than a vague sentence.
Show it one example
Describing a tone of voice is hard. Pointing at one is easy. Paste a paragraph you like and say so. Professionals call this few-shot prompting. I call it showing the apprentice what you meant.
Here is a product description I like :- "Hand-thrown mug, holds a proper cup of tea, survives the dishwasher and the children. Made in Oxfordshire." Write five more in the same voice and length for the items below. Keep the dry humour. No exclamation marks. [paste items]
Set the boundaries
Limits sharpen output. Word counts, banned phrases, spelling variant, reading age, what to do when it does not know. Four short rules often do more than a paragraph of encouragement.
Write a blog intro about our farm shop.
Write a blog intro about our farm shop in 80 words. British English. No exclamation marks, no 'nestled', no 'hidden gem'. If you need a fact I have not given you, leave a [GAP] marker instead of inventing one.
The [GAP] trick is the one I use most. It turns made-up facts into a to-do list.
Make it interview you first
When you are not sure what the AI needs, let it ask. This turns a one-shot guess into a short conversation and usually saves a round of rewrites.
I need a one-page proposal for a new client. Before you write anything, ask me up to five questions you need answered. Ask them one at a time and wait for my reply to each. Once you have enough, say so and write the draft.
Split big jobs into steps
Long, tangled requests produce long, tangled answers. A proposal is really four small jobs wearing a coat. Do them one at a time, checking each before you move on.
- BriefAsk the AI to restate the job in its own words so you can spot a misunderstanding early.
- OutlineAsk for headings only. Move things around before any prose exists.
- DraftAsk for one section at a time, pasting your approved outline back in.
- TightenAsk for a 20 per cent cut and a plain-English pass.
Make it check its own work
AI is a decent editor of its own first draft, provided you give it a hard-nosed reviewer to impersonate. This will not catch every error, so you still read it. It will catch the lazy ones.
Act as a sceptical finance director. Read the draft above and list the three weakest claims, the two things a reader will not believe, and any figure that needs a source. Do not rewrite it. Just list the problems.
For anything factual, see how to judge the accuracy of AI information. A confident tone is not evidence.
Keep a prompt library
The tenth tip is the one that compounds. Every time a prompt works, save it. Within a month you have a small toolbox and you stop starting from a blank box. A shared document is enough.
| Column | What goes in it | Example |
|---|---|---|
| Name | What the prompt does, in plain words | Late payment email, firm |
| Prompt | The full text with [brackets] for the bits that change | You are a credit controller... |
| Works in | The tool and mode you tested it in | ChatGPT, thinking mode |
| Last tested | A date, because tools change | 17 Oct 2026 |
Most tools now let you store this as saved instructions or a project. Context management is the grown-up name for the same idea, and it is where prompting gets properly powerful.
Which tip should I use first?
Start with the symptom, not the list.
| If the output is... | Try tip | Why |
|---|---|---|
| Generic and bland | 1, 5 | Give it a role and an example to copy |
| Right topic, wrong angle | 2, 7 | State the purpose, or let it ask you |
| Full of invented facts | 3, 6 | Paste the facts and allow a [GAP] marker |
| Long and rambling | 4, 6 | Fix the shape and the word count |
| Fine but not quite there | 9 | Make a reviewer attack it |
| Different every time | 10 | Save the winner and reuse it |
What mistakes do people make with prompts?
- Writing a sentence when a paragraph was needed. Short prompts get short thinking.
- Rewriting the whole prompt after one bad answer. Reply with what was wrong. The AI remembers the conversation.
- Treating the first answer as final. Second drafts are where the value is. Say "shorter", "plainer" or "more like the example".
- Pasting sensitive data into a tool nobody approved. Fix the policy before you fix the prompt. Our guide on stopping staff pasting confidential data into ChatGPT is a good place to start.
Prompt engineering: questions people ask
What is the most important prompt engineering tip?
Paste the real material. An AI that works from your notes, emails or data will beat the cleverest instruction wrapped around nothing. Role, format and limits all help, but context does most of the heavy lifting.
Do I need to learn special prompt syntax?
No. Modern tools understand plain English. The skill is in being specific about audience, purpose and format, not in memorising formulas.
Some tools respond well to structure, such as headings or labelled sections inside a long prompt. That is tidiness, not magic.
Do these tips work in ChatGPT, Claude, Copilot and Gemini?
Yes. All ten work across the major assistants because they describe good briefing, not one product's quirks. Features differ. For tool-specific advice see our guides to ChatGPT and Microsoft Copilot.
How long should a prompt be?
As long as a good brief to a colleague would be. For a simple question, one line. For a document, a paragraph or two plus the material. Length is fine when every sentence changes the answer.
Is prompt engineering a real job?
It was hyped as one. In practice it has become a skill inside other jobs, the way spreadsheets did. The people who benefit are accountants, marketers, managers and owners who write clear instructions.
Where can I learn prompt engineering properly?
Practise on real work for a few weeks, then consider structured training. Our Prompts, Context and Skills deep dive and Basic AI Training are built for exactly this.
Keep reading
- How to write prompts for AIThe full method, with the Five Rivets framework and a worked example.
- How to use ChatGPT effectivelySettings, habits and features that change the quality of every answer.
- How to use Claude effectivelyWhat Claude does differently and how to brief it.
- AI at work cheat sheetOne page to pin next to your monitor.
Practise these on your own work, with someone watching
Reading about prompts is like reading about swimming. Our half-day deep dive has your team writing, testing and saving prompts for their real tasks, in plain English.
- Basic AI Training Start here if AI still feels a bit slippery.
- AI Basics Deep Dive Go further once the basics are second nature.
- AI Capability Assessment Find out where your team really stands.
Harry Lang
Harry Lang founded The Oxford AI School in Witney, Oxfordshire. He trains UK businesses and teams to use AI well, writes the Marketing Unfiltered newsletter, and is the author of Learn To Fly With AI.