What is a prompt and how to write one: a formula for strong requests

In short: a prompt is the text in which you set a task for an AI in plain words. Answer quality depends 80% on prompt quality. A working formula is Role + Task + Context + Format: say who the model should be, what to do, under what conditions, and in what shape to return the result. Below: the formula, before/after examples and ready templates.

What a prompt actually is

A prompt (from "to prompt" — to nudge, to cue) is your message to the AI: a question, an instruction or a task. The model can't read your mind or see the context of your life — it works only with the text you gave it. Hence the core principle: the sharper the input, the more useful the output. If you're just starting, first read the beginner's guide to ChatGPT — it covers sign-up and first steps.

Why a one-liner almost always yields a weak answer

The request "write a post about coffee" is taken literally and averaged out: you get a faceless text for no one. The model doesn't know your audience, the tone, the length or the goal. It has to guess all of it — and it guesses from the "average internet". The result is filler.

Think before reading on: if you handed the same task to a real person who'd never done it, what would you make sure to explain? That's exactly what a short prompt is missing.

The Role-Task-Context-Format formula

• Role — who the model should act as: "You are an experienced editor", "You are an employment lawyer". The role sets vocabulary, tone and depth.
• Task — what exactly to do, one clear verb: check, rewrite, compare, draft a plan.
• Context — the inputs: for whom, why, constraints, facts. Paste your own text or data here too.
• Format — the shape of the output: a 5-item list, a table, an email, a short paragraph.

Before and after

Before: "write a post about coffee".

You are a content marketer for a coffee shop next to an office centre.
Task: write an announcement post for the autumn menu.
Context: audience — office workers aged 25–40,
friendly tone, no hype; new items — pumpkin latte and filter coffee.
Format: 4–5 short paragraphs + a call to drop by for lunch, no hashtags.

The second prompt yields publish-ready text because it leaves no gaps for the model to fill at random.

Five techniques that strengthen any prompt

1. Give an example of what you want. "Here's a sample style: …" — the model matches it more precisely than any description.
2. Ask it to think step by step. For reasoning tasks: "Work through it step by step, then give the conclusion."
3. Add constraints. "No longer than 150 words", "only from the attached text" — limits raise quality.
4. Permit not knowing. "If data is missing, say so — don't invent." This lowers the risk of false facts.
5. Iterate. The first answer is a draft: "shorter", "different tone", "replace the example" — refine in dialogue.

Common beginner myths

• "There's one secret prompt for everything." No: the formula is one, but you supply fresh context each time.
• "Longer is always better." What matters isn't length but having the needed inputs without clutter.
• "Politeness makes the model smarter." "Please" doesn't hurt, but role, context and format decide quality.

Do it now (2 minutes, on paper)

Take any task of yours and lay it out across four slots: Role / Task / Context / Format. Even without an open chat you'll already see which inputs were missing. The skill transfers across all models — as the comparison of ChatGPT, Claude and Gemini shows. And why even a perfect prompt doesn't let you trust facts blindly is in our article on AI hallucinations.