What is a prompt? The complete beginner's guide

Ask an AI "write an email to follow up with a client" and you'll get a polite, correct, unusable text. The tool knows nothing about your client, the amount owed, the invoice date, or the way you usually talk. Add four lines and the same tool gives you an email that only needs a read-through. Those lines are part of your prompt, and the prompt is the only thing you truly control when you use an AI.

A prompt is the message you send to the AI

The word comes from the verb to prompt, to nudge into action, and it has a meaning in the theater too: the prompter is the person who feeds the actor a line. In computing, it already referred to the prompt waiting for your command in a terminal. With AI tools, it has taken on a broader meaning: everything you give the machine to get a response.

In ChatGPT, Claude, Gemini, or Le Chat, it's what you type in the text box. In an image generator, it's the description of the image you want. A prompt can be a ten-word question or run a full page with a document pasted into it. You'll also come across the words "query" and "request," which mean the same thing.

There's no syntax to learn, and no magic formula. Plain language is enough. The skill resembles that of someone handing a job to a new colleague: this colleague is quick and willing, but arrived this morning and knows nothing about your situation. What you tell them in the first thirty seconds decides what they hand back.

What happens when you press Enter

A language model doesn't go looking for a ready-made answer in a database. It was trained on enormous amounts of text, and it writes its answer piece by piece, calculating at each step what is most likely to come next, given what you wrote. If the mechanism interests you, this guide explains it without jargon.

To write good prompts, remember four consequences:

• What you don't write, the AI fills in. It picks the most probable option, and therefore the most average one. That's why vague requests produce generic texts.
• It can be wrong with confidence. A well-turned answer isn't necessarily accurate. Style says nothing about reliability.
• It only sees what you give it. It knows nothing about your company or your file unless you say so, or unless the tool has a memory or access to your files that you've turned on. These features vary between tools and plans: at the time of writing, the simplest thing is to check in the tool.
• It doesn't answer the same way twice. The same prompt can produce two different texts from one try to the next. That's normal, and it matters as soon as you want to rely on a prompt.

Questions people ask at the start

Do I need to use a special language? No. The major tools understand plain, everyday language very well, and you'll describe your situation better in your own words. If you want an answer in another language, just ask.

Do I need to be polite? Do as you normally would; it doesn't replace anything. "Please" won't make up for missing information, and a curt but precise instruction will give better results than a friendly but fuzzy request.

Does a prompt work everywhere? A clear text prompt carries over fairly well from one tool to another, even if the answers differ. Image generators are a case apart: each has its own habits and, sometimes, its own parameters.

From a vague prompt to one that works

Take a restaurant owner who wants to fill her dining room at lunchtime. Here's what people write spontaneously:

Write a post for my restaurant.

You can guess what follows: a cheerful text about "authentic" cooking prepared "with passion," which could apply to any restaurant in the country. The AI has nothing to hold on to, so it produces the average of all the restaurant posts it may have read.

Here's the same request, written the way you'd hand it to a person:

Write a Facebook post for my restaurant, [NAME], a neighborhood bistro in [CITY].

Goal: bring in customers at weekday lunchtime for our new set menu at €[PRICE].
Readers: mostly people who work in the neighborhood and have an hour for lunch.
Dishes of the week: [DISHES].

Tone: simple and warm, like an owner talking to her regulars. No superlatives, no emojis.
Length: 60 words maximum.
Last sentence: the service hours ([HOURS]) and the number to book ([PHONE]).

The prompt is longer, but length was never the goal. Look instead at what changed:

• The channel is specific, and a Facebook post isn't written like a poster.
• The goal tells the AI which argument to put forward.
• The readers point toward quick service rather than a romantic dinner.
• The real facts (the dishes, the price) no longer need to be invented.
• The limits on tone and length prevent bloated text.

Each addition takes a decision away from the machine and hands it back to you. You'll find the detail of these ingredients in our article on the 6 elements of a good prompt.

The main families of requests

Not all prompts are equal when it comes to the risk of error. This table helps you know what you're getting into.

The practical rule: the more of the material you supply, the safer the result. To get started, begin with the requests at the top of the table, on texts you know well. You'll see right away what the AI does with your work.

Three prompts to copy to get started

These three prompts cover common needs. Replace what's in brackets and keep the rest.

Making a text clearer

Useful for a delicate email, meeting minutes, or a message to a team.

Here is a text I wrote for [RECIPIENT]:

[PASTE YOUR TEXT]

Rewrite it to be clearer and shorter, without adding any information and without changing the meaning. Keep my level of politeness.

Then list the three most important changes you made, one line each, so I can accept or reject them.

The last instruction saves you time: you see right away where the AI touched your text.

Understanding a complicated document

Imagine an association volunteer who receives a twelve-page grant agreement. Before pasting anything in, remove names, contact details, and sensitive information, and check what you're better off never pasting into a chatbot.

I'm [YOUR ROLE] and I have no training in this field. I need to understand the document below.

[PASTE THE DOCUMENT]

1. Explain in five simple sentences what it's about.
2. List what I need to do, with the dates if they're given.
3. List the unclear points, or those I should have confirmed by a qualified person.

Rely only on the text provided. If a piece of information isn't in it, write "not specified" instead of deducing it.

The last sentence reduces the risk of invention without removing it. Check every date and every amount against the original.

Having the AI interview you

When you don't know what to specify, reverse the roles. This is probably the most useful prompt for a beginner.

I want [WHAT YOU WANT TO GET].

Before answering, ask me the questions you need to do a job suited to my situation. Five questions maximum, one at a time, waiting for my answer between each.

When you have enough to go on, give your proposal and state the assumptions you made in my place.

A craftsman who wants to write his company's introduction will get, this way, a list of points he wouldn't have thought of, and will learn along the way what the AI needs to know for next time.

A prompt carries on in the conversation

Being a beginner often means believing you have to get everything right the first time. In most chat tools, the answer stays in the exchange and you can correct it: "keep the second paragraph and delete the rest," "you assumed the client was an individual, it's a company." These messages are prompts just like the first one.

The most effective approach is to say precisely what's wrong. "Redo it" or "that's bad" gives the AI nothing to go on, and it will try something else at random. The full method is in our article on improving a prompt through iteration.

When a prompt ends up working well, keep it. The day you have the same task to do again, you'll only need to change the information in brackets. If you'd rather start from work that's already been tested, the YouPrompt catalog shows an excerpt and a real result before any purchase.

What a good prompt will never do for you

A well-crafted prompt clearly improves answers, but it doesn't make them reliable. Keep these limits in mind:

• It doesn't eliminate errors. The AI can invent a reference, a legal article, or a figure, even when you ask it not to. Every result must be reread, and the facts checked against a source.
• It doesn't replace your knowledge of the subject. You'll spot an error in your own trade, much less so in a field you don't know. On those topics, be doubly careful.
• It doesn't protect your data. What you paste into a tool may be stored or used according to its terms and your settings. Read them before putting anything confidential in. The CNIL also publishes useful guidance on these topics.
• It doesn't relieve you of deciding. On an important medical, legal, or financial question, the AI can help you prepare your questions. The answer that commits you comes from a professional.

Your first try, in ten minutes

Choose a writing task you did this week and whose subject you know well. Write your request the way it comes to you, send it, keep the answer. Then open a new conversation and rewrite the request as if you were handing it to a colleague who arrived this morning: for whom, to what end, with which facts, in what form, how long. Compare the two answers side by side and note the sentence in your second prompt that changed the result the most. That's the one you'll need to remember to write next time.