How do you write a good AI prompt? Leave nothing to guess

Why does AI work for my colleague and not for me?

Your colleague uses the same tool as you. The difference is in the request.

When the AI gets something wrong, I give my learners a deliberately extreme answer: consider that the mistake is yours. Not to blame you. You sign what goes out, and "the AI said so" is not an answer a client will accept. And your request is the only part of the exchange you control.

On your side, three things make the AI fail:

• An impossible request. Ask the AI to "build a shoe": the AI will draw a shoe, describe a shoe or refuse, and none of the three answers will suit you.
• A vague or incomplete request. The AI fills in everything you left out, with its own choices.
• No check. The answer goes out as the AI wrote it. How much to check, and how, is in the guide Why does AI make things up?

This guide is about the request. See AI as an office tool, like Excel: nobody blames Excel for a formula they typed wrong. The habit to drop is hoping: typing a request in thirty seconds, pressing Enter, and expecting a great answer to a request you could not have explained yourself. The AI cannot carry out a task that you could not explain.

Invest at the start instead. Framing the first request well takes a few minutes; correcting a vague request takes ten messages, and every answer you have to redo is time spent twice.

Do you need a prompt formula (role, context, task, format)?

No. A prompt formula is a checklist for you, not a magic spell for the AI.

Online, you will find formulas such as "role, context, task, constraints, format". A formula helps you not to forget anything, and that is all a formula does. The AI does not need your request in that order, or under those headings.

The same goes for "act as an expert recruiter": giving the AI a role changes very little in an ordinary request. A role is useful in one case: when you build a specialized assistant that you will reuse, such as a Gem in Gemini, a custom GPT in ChatGPT or the instructions of a Claude project.

Two qualities make the difference, whatever the form. The request is explicit: the AI knows exactly what result you expect. The request is complete: the AI has every piece of information the task needs. The next two sections take the two qualities one at a time.

How do you make an AI request explicit?

Replace every adjective with a criterion you could tick off.

"Short", "fast", "professional", "reliable", "perfect": for the AI, these words mean nothing precise. Is a short email 10 words or 100? The AI chooses for you. "Fast" means three minutes for cooking pasta and several months for one company buying another. "Under 80 words" leaves the AI nothing to choose, and lets you check the answer in two seconds.

The test: for each thing you ask for, could you tick a box to say it was done? If you cannot judge the answer, the AI cannot either.

A rule becomes checkable when you write it as a condition: not "the stock must stay reliable", but "when a product reaches its expiry date, its status changes to expired".

Choose the action verb carefully

The verb tells the AI what kind of result to produce. A single word can turn a two-page document into a thirty-page one.

Avoid "do" and "make": "make an email" can mean writing the text, creating a draft in your mailbox or sending it. "Write the text of an email" leaves a single meaning.

Choose words with a single meaning

For the AI, "bug" points to software, while "anomaly" can just as well mean a power cut in the street. The effect of one word is small. But every vague word leaves a small gap, and the small gaps add up.

How do you give the AI all the context it needs?

Use the colleague test: could a colleague who knows nothing of your situation do the task with your message alone, without asking you a single question?

Picture a colleague from another team, not the one who sits next to you and already knows your clients. Wherever that colleague would need to ask you something, the AI will guess instead. That is the golden rule at work: what you don't give the AI, the AI decides for you.

Context even decides which answer is the right one. "How tall is the Eiffel Tower?" asked during a timed quiz calls for one number, nothing more. The same question from someone curious about the monuments of Paris calls for the history of the tower's height. The words are identical; nobody but you knows which answer you need, so you have to say it.

Take the request "Write a short email to Mr. Dupont to remind him about the invoice." You have known Paul Dupont for fifteen years, you are on first-name terms, and you have lunch together every month. None of that is in the request, so the AI writes "Dear Mr. Dupont", in the tone of a stranger.

Show the AI an example instead of describing the result

You could describe a picture for four hours and never get the picture redrawn faithfully; show the picture, and the AI sees it. The same goes for a style, a layout or a tone: give the AI examples of the result you expect. Three to five examples are often enough for the AI to see the pattern. Leave out examples that are off topic: off-topic examples blur the pattern.

A template to fill in works the same way: give the AI your usual document with the boxes left empty. And for long documents, a short excerpt is enough to show your style: there is no need to send five whole files.

Complete does not mean everything

Send what the task needs, and nothing more. Sending a whole dictionary to get three definitions buries the three words you care about. To draft a reply, the AI needs neither the client's phone number nor the client's email signature: write [client] instead of the name, and put the name back yourself. Extra data only adds noise to the answer. More in the FAQ: Can I paste client data into AI?

Read the guide: does AI remember what you told it?

How do you get the AI to ask you questions before it starts?

Add one sentence at the end of your request. The AI then asks you about every point the AI would otherwise have decided alone.

Give a precise number. With "ask me questions if needed" or "if you don't understand", the AI usually answers that it has understood everything. Each question the AI asks you is a choice the AI would have made for you.

Write your best request first, then ask for the questions

The questions come after your best request, not instead of it. "Write a short, professional email. Ask me 10 questions" puts you back to hoping. Write everything you know first; the questions then find what is still missing.

How many questions should you ask the AI for?

As many as the stakes and the difficulty call for. 10 is a start for a simple task; for real work, I usually ask for 20 to 30. Answer for real, in detail, never with a bare "yes", "no" or "maybe": your answers are the context the AI was missing. And raise the number with the stakes: if a single mistake could cost you a fine, ask for many more than 10.

When should you stop the AI's questions?

When the last questions repeat earlier ones, or become trivial. If every question is still relevant, ask the AI for five more.

Why doesn't "improve my prompt" work?

Because the AI learns nothing new about you. The AI rewrites your request by guessing what you meant, then carries out the rewritten request by guessing again.

Your request has gaps you cannot see, because you know your situation too well. "Improve my prompt" asks the AI to fill those gaps without any new information. The AI fills the gaps with plausible guesses, and the polished wording makes the guesses harder to spot. One layer of interpretation becomes two.

Two more traps. "Improve" is itself a vague word: improve compared with what? And the rewritten request is longer than yours, so most people send it without reading it, and control the result even less.

Put the questions first, and the rewrite second:

What are the four levels of an AI request?

Four ways to ask, from hoping to planning. Choose the level according to the stakes.

Each level adds to the previous one: a level 3 request is still explicit and complete. At level 4, even after all that work, rereading the plan usually turns up something to correct, and that correction is the whole point of the plan. And there is no need to bring out level 4 to ask the height of the Eiffel Tower.

What should you check before you send a request to the AI?

Five questions to ask yourself before you press Enter.

• Could a colleague who knows nothing of my situation do the task without asking me a single question?
• Could I tick off each thing I ask for? No "short", no "professional": numbers and facts.
• Does the verb name the result I want: summarize, synthesize or detail?
• Have I shown an example of the result, if I have one?
• If the task matters: does my request end with a precise number of questions?

A good request leaves the AI fewer blanks to fill. A good request does not replace checking: the guide Why does AI make things up? shows how much to check, and how.