Write an AI prompt that gets a useful answer

Google trained all of us to strip a question down to keywords, and that habit is the single biggest thing standing between a beginner and a decent answer. A model can only work with what you put in front of it, and three words put almost nothing there. The five moves below close that gap, and the template that bundles them sits right after.

To get a useful answer from ChatGPT or Claude, give it a short brief rather than a search query: the task, the context behind it, the format you want back, and one example of good. Anthropic's own guidance is to treat the model like a capable new employee who has none of your background.

Five moves that turn a search into a brief

1. Say what you want in plain words. Anthropic's own prompting guidance puts this first, and its test is the useful part: show your prompt to a colleague who knows nothing about the task and ask them to follow it. If they would be confused, the model will be too. A human reader catches the missing bit faster than you will staring at your own sentence.
2. Say why, not only what. This is the move that changes the most for the least effort. Anthropic's docs use a before and after: "NEVER use ellipses" works less well than "Your response will be read aloud by a text-to-speech engine, so never use ellipses since the text-to-speech engine will not know how to pronounce them." Give the reason and the model generalises from it, so it also handles the twenty related cases you did not think to list.
3. Ask for the shape of the answer. Say how long, in what form, for whom. Describe the target rather than banning what you do not want: Anthropic's own example swaps "Do not use markdown in your response" for "Your response should be composed of smoothly flowing prose paragraphs." A described target gives the model something to aim at, where a ban only tells it where not to stand. Vanderbilt teaches the useful flip on this: describe who is reading rather than who the model should be. Their example is "Explain large language models to me. Assume I am a second grader who gets bored easily."
4. Show one example of good. Paste a thing you liked, from you or from anywhere, and say what you liked about it. Anthropic suggests three to five examples for best results, though one beats none by a mile. Every adjective you could reach for is doing guesswork that a real sample settles in a second.
5. Make it ask you questions first. Add one line: "Before you answer, ask me any questions you need to give a genuinely good answer." Vanderbilt University teaches this as the Flipped Interaction Pattern, phrased as "I would like you to ask me questions to achieve X. You should ask questions until condition Y is met". It hands the hard part back to the tool, which is fair, because the tool is the one that knows what it is missing.

The Brief

Here is the template. Fill in the lines that matter for your ask and delete the rest, then paste it above your question. It works the same in Claude and in ChatGPT.

ROLE (optional)
[Answer as the kind of person whose judgement fits this job, for example "someone who has written a hundred of these", "a plain-speaking editor who cuts hard", "a careful accountant who never rounds". This shapes the tone and register only; it will not make the answer more accurate, so keep it light and skip it if it does not obviously help.]

TASK
[What you want, in one plain sentence, the way you would ask a competent colleague across the desk. Name the finished thing you want in your hands, the actual deliverable, like "write the four-line message I send my landlord to get a repair date for the boiler" rather than a vague "help with my landlord". Name the format in the same breath if you can.]

WHY I NEED IT, AND WHO READS IT
[The reason behind the task and who is on the other end, in two or three real sentences. This is the layer that kills generic answers, because it lets the model handle the cases you never thought to spell out. Say what is at stake and how the reader is likely feeling. For example: "This goes to existing customers, some of whom will feel stung by the price rise and some of whom barely remember signing up, and I want them to feel told straight, rather than sold to."]

WHAT I ALREADY KNOW
[Everything the model cannot see and would otherwise guess at: the situation, the constraints, the history, the real names and the real numbers. Paste it in raw rather than tidying it into a summary, because the specifics are the whole reason the answer comes out as yours and not as anyone's. If a number matters, give the real one now.]

FORMAT I WANT BACK
[How long, in what shape, and for whom, described as a target to aim at. "About 120 words, first person, plain paragraphs, no headings, written for someone who has never met the project and has thirty seconds" gives the model something to hit. A pile of "do not do this" only tells it where not to stand.]

AN EXAMPLE OF GOOD
[Paste one real thing you liked, from you or from anywhere, and say in a line what you liked about it, for example "I liked that this one said the awkward part in the very first sentence and then got practical". One real sample settles in a second what ten adjectives only gesture at. If you genuinely have nothing to paste, say so and skip this line rather than inventing a style.]

WHAT TO AVOID
[Two or three hard no-gos only, the lines you would strike out on sight: "no exclamation marks", "never promise a delivery date", "no American spelling", "do not call it a solution". The format line above already carries most of this, so keep this short.]

WHAT TO GIVE ME BACK
[The finished thing, plus the one or two things I will reach for the second I have it, so I am not straight back asking. For example: "the email, then a two-line version for a text, then the two replies I am most likely to get and how to handle each". For a decision rather than a draft: "your pick in the first line, the strongest reason for it and the single strongest reason against, and what I should watch in the first month to know if it was right".]

BEFORE YOU ANSWER
First, work out what I am actually asking for and settle the approach before you draft. Do not jump straight to writing. Then, if anything you need is missing or could sensibly go two ways, ask me up to three questions and wait for my answers. If you already have everything, answer straight away. Never invent a fact, a name, a number, or a source to fill a gap: put it in [square brackets] for me and, at the end, tell me plainly what you left out. When you finish, name the single part you are least sure about, so I know where to look first.

And here is the whole thing filled in once, so you can see it working rather than picture it. Copy this, swap in your own lines, and delete any you do not need.

ROLE: Answer as a plain-speaking editor who cuts hard and hates a warm-up.

TASK: Write the LinkedIn post announcing that our prices are going up from 1 March.

WHY I NEED IT, AND WHO READS IT: Existing customers will see this. Some will feel stung, some barely remember signing up. I want them to feel told straight, rather than sold to, and I want the awkward part said in the first line so nobody feels ambushed halfway down.

WHAT I ALREADY KNOW: Prices rise on 1 March by about 8 per cent. Existing customers keep their current price until their next renewal. We make a booking app for hair and beauty salons, so my reader is a busy salon owner.

FORMAT I WANT BACK: About 120 words, first person, plain paragraphs, no hashtags, written for someone reading on their phone between clients.

AN EXAMPLE OF GOOD: [paste a post you admired here] I liked that it led with the change and did not bury it under three lines of thanks first.

WHAT TO AVOID: No exclamation marks, no "we are excited to announce", no corporate softening of the actual news.

WHAT TO GIVE ME BACK: The post, then a two-line version I can send as a text to my regulars, then the two objections I am most likely to get in the comments and a calm reply to each.

BEFORE YOU ANSWER: Work out the shape before you draft, then ask me anything you still need and wait. Do not invent a figure or a date I did not give you. At the end, tell me the line you are least sure about.

Forty seconds of filling that in, and the answer stops sounding like a press release. Here is roughly how the reply itself shifts on that pricing post, illustrative rather than a captured transcript:

Same model, two minutes apart. The difference is the first line, and the first line is the part people actually read.

Here is a fuzzier one, a decision rather than a draft, where the context and the questions do the work:

And the other direction. For "what is the capital of Portugal", use none of this. A short question is the right tool for a simple fact, and The Brief is for the asks where good is hard to describe. Keep as many lines as the job needs and bin the rest.

When the first answer misses

It usually will, somewhere. Say which bit and what you wanted instead: "the second paragraph is too formal, rewrite that one the way you wrote the first." Correcting in place keeps everything you already gave it. Opening a fresh chat throws all of that away and you type the brief again for nothing.

Where the line is

The persona study: When "A Helpful Assistant" Is Not Really Helpful, Zheng and colleagues, arXiv 2311.10054.

One more tool worth knowing

If you would rather fill in boxes than a template, a developer called Mahmoud Zalt built a free AI Prompt Builder that does the same job as a form: you fill in role, task, context, output format, tone and constraints, watch the prompt assemble live, and copy it. There is no signup and no usage limit, and the page states that it runs entirely in your browser with nothing transmitted, stored or logged, which matters if your context is not something you want sitting on someone's server.

Free, no account: zalt.me/tools/prompt-builder. It is one developer's free tool and he sells consulting from the same site, so take it as a handy thing he made rather than a neutral utility.

Try it on the next thing you ask

Do not rewrite your habits today. Take the next question you were going to type, paste The Brief above it, and fill in two lines: why you need it, and what good looks like. Those two alone change most answers. Keep the template in a note on your phone and it takes about forty seconds a time after that. Once it has earned its keep, paste it into a Claude Project's instructions and every chat in that project starts already briefed.

FAQ

How do I write a better AI prompt?

Give it a brief rather than a search query. Say what you want in plain words, say why you need it and who reads it, describe the format you want back, and paste one example of good. Then add a line telling it to ask you questions before it answers.

Does telling AI to act as an expert actually work?

Less than you have been told. When researchers tested this properly across many roles and models, adding a persona to the system prompt did not improve factual accuracy. Roles help set tone. Context, constraints and examples are what improve the answer.

Do I need a long prompt every time?

No. A short question is the right tool for a simple factual ask. The brief earns its place when the request is fuzzy, when good is hard to describe, or when you have already had answers you did not want.

Do the tipping and "this is important to my career" tricks work?

They were a workaround for older models. Anthropic's current guidance is to dial back aggressive phrasing like "CRITICAL: You MUST", because newer models over-react to it. Spend the line on context instead.