How to Write AI Prompts: A Beginner's Guide

What a prompt is and why it decides everything

A prompt is your request to an AI: the text you send to ChatGPT, Midjourney or any other model to get a result. And here is the key thing to grasp: the model does not read minds and knows nothing about the context of your life. It works purely with what you typed in that box. Anything you did not say, it will either fill in its own way or ignore.

Why does understanding the mechanics matter? Take a look at our breakdown of how ChatGPT works: the model predicts the continuation of your text. That means your prompt is literally the beginning it continues. A vague beginning yields a vague continuation. A clear, detail-rich beginning yields a precise, useful answer. A prompt is not a "wish" — it is input data that directly shapes the result.

The anatomy of a good prompt

A strong prompt is almost always assembled from a few "building blocks." Not every request needs them all, but the harder the task, the more of them you should add. Here is what a quality request is built from:

Let's look closer at the three most underrated blocks:

• Specifics over generalities. "Write about sports" → "Write 5 tips for a beginner on how to start running without injury." A vague request breeds a vague answer.
• Role. The phrase "You are an experienced pediatrician" tunes the model to the right body of knowledge, vocabulary and tone. It is a simple, powerful trick: one line noticeably changes the character of the answer.
• Constraints. Limits cut out the guesswork: "keep it under 150 words," "no intro," "plain language, no jargon," "do not invent facts." The tighter the limits, the more predictable the result.

A before-and-after example

Theory comes alive in contrast. Take a typical task — an email to a customer — and see how the same building blocks turn an empty request into a working one.

See the difference? The second request contains a role (support manager), context (3-day delay, courier's fault), a task (apologize and compensate), a tone (warm, no corporate-speak) and a length limit (120 words). The model got a full brief — and the answer will be right the first time, no rewriting.

Examples in a prompt: the power of few-shot

There is one trick that lifts quality more than any other — few-shot ("a few examples"). The idea is simple: instead of describing the desired result in words, you show the model 2–3 ready samples. Examples steer it more precisely than the most detailed instructions.

Let's sort out the terms so nothing gets confusing:

• Zero-shot ("zero examples") — you just describe the task in words, no samples. Fine for simple requests.
• Few-shot ("a few examples") — you embed a couple of examples of format and style in the prompt. Indispensable for unusual or finicky tasks.

A few-shot example for generating short slogans: "Come up with a slogan in this style. Coffee shop → 'Mornings start not with coffee, but with us.' Bookstore → 'Thousands of lives on one shelf.' Now: flower shop → …". By showing two samples you set the length, the intonation and the device — and the third slogan will come out in the same key. This works for images too: in image generators, specifying style, lighting and angle plays the same role that examples play in text.

Iteration: a dialogue, not a command

The main beginner mistake is treating a prompt as a one-shot command: ask once, dislike it, close. In reality talking to an AI is a dialogue. The first answer is a draft that can almost always be improved with one or two edits, without starting over.

This dialogue-based approach saves time and almost always beats trying to "guess the perfect prompt on the first go." Two or three clarifying turns, and a mediocre answer becomes exactly what you need.

Common beginner mistakes

Finally, the rakes people step on most often:

• Too general a request. "Help me with some text" — the model has no idea which text or how. Always add specifics.
• Everything in one heap. Five different tasks in one prompt confuse the model. Break the complex into steps.
• No format or length. Leave it out and you get a random structure. Ask for a list, a table, a word count.
• Blind trust. The model is confidently wrong sometimes (that is "hallucination," see our breakdown of how ChatGPT works). Double-check facts, numbers and quotes.
• Politeness over clarity. "Please" means nothing to the model. Spend words on substance: what, for whom, in what form.

You now have a working framework: specifics, context, role, examples, format, constraints — plus the habit of iterating. This is the very skill that separates "AI is useless" from "AI saves me hours." Want to understand more deeply what happens on the other side of the request? Read about how a neural network is built and generative AI.