Writing better prompts: 7 key principles

Most people type a question into an AI assistant the way they would type it into Google. The result: a vague answer that circles the topic without ever landing on it. The problem rarely comes from the tool. It comes from the instruction.

Here are seven principles you can apply from your very next conversation. None requires technical skill.

1. Give a role and a context

A model answers better when it knows who it is writing for. “Explain the taxation of the self-employed to me” gives a lecture. “I am a freelance graphic designer, just starting out this year in France. Explain what I should set aside for taxes, in plain language” gives a usable answer.

Context is not a luxury: it is what turns a generic answer into an answer for you.

2. Say what format you expect

Do you want a list? A table? A three-line paragraph? Say so. Without instruction, the model chooses for you, and it often chooses long.

“Summarise in five bullets, one sentence each” will save you considerable time on content you need to skim.

3. Show an example

If you expect a precise tone or structure, show one. Paste an email you find well worded and ask “write me a reply in this style.” A single example guides the model far better than a long abstract description.

4. Break down complex tasks

A request that contains five sub-questions produces an answer that skims five. Handle them one by one. You can chain them: “Let’s start with the outline. Once it suits me, we’ll write.”

This step-by-step approach almost always gives a better result than a sprawling instruction. It is also the starting point for automating a repetitive task, step by step.

5. Let the model think

For a piece of reasoning or a calculation, explicitly ask for the steps: “Detail your reasoning before concluding.” A model that lays out its path is wrong less often, and you can more easily spot where it slips.

6. Correct instead of restarting

The first answer is a draft, not a verdict. Rather than starting from scratch, adjust: “Too formal, make it more direct,” “You forgot the budget, include it.” The conversation is your best precision tool.

7. Check what matters

An AI assistant sometimes states false information with the same confidence as a true one. For a figure, a date, a legal reference, verify at the source. Use the tool to go fast, not to excuse yourself from judging.

The skeleton of a good prompt

The seven principles fit into a simple structure worth keeping at hand. Five blocks are enough:

Role: you are [an accountant addressing a self-employed person].
Context: [I am starting my activity this year, in France, on my own].
Task: [explain what I should set aside for taxes].
Format: [a list of 5 points, one sentence each, plain language].
Constraints: [no jargon; if a figure depends on my status, say so].

You do not have to fill all five blocks every time. But when an answer disappoints, it is almost always because one of them is missing — most often the context or the format.

Before / after: a prompt transformed

The same need, phrased twice:

• Before: “Write an email to follow up with a client.” → A generic result, random tone, unpredictable length.
• After: “You are a salesperson. The client hasn’t replied to my quote sent ten days ago. Write a short follow-up (5 lines max), cordial but nudging for a reply, offering a call slot this week.” → A directly usable email.

The difference is not the model’s talent: it is the role, the context and the format you gave it.

When the answer misfires: the principle to reach for

Each common flaw maps to a precise principle. This table works as a quick fix:

The symptom	The principle to apply
Answer too long, indigestible	2 — impose a format (“5 bullets, one sentence”)
Generic answer, “for everyone”	1 — give a role and a context
The tone doesn’t fit	3 — show an example
The AI skims a complex request	4 — break it into steps
A piece of reasoning or a calculation looks wrong	5 — require the reasoning steps
Almost right, but a detail is missing	6 — correct within the conversation
A figure or a date looks doubtful	7 — verify at the source

Key takeaway

A good prompt fits into one common-sense sentence: say who you are, what you want, in what form, and correct along the way. These habits apply to every assistant on the market. They do not make AI perfect, but they make the difference between an answer you throw away and one you use. To apply them without starting from scratch, pick from our 10 ready-to-use prompts for the office. And for other step-by-step methods, browse our practical guides.