FR Newsletter

AI for Work

Write better prompts with a simple six-part method

When an AI assistant gives a vague answer, the problem is often the question. The official prompting guides of Anthropic, OpenAI and Google say much the same thing in different words: be clear, give context, show what you expect, and refine. Here is that advice turned into a six-part method you can use every day.

Why a prompt fails

An assistant only knows what you write. It does not know your job, your audience, or the result you have in mind. Anthropic sums it up with an image: think of the model as "a brilliant but new employee who lacks context on your norms and workflows". Its golden rule is a good test for any prompt: would a colleague with no background on the task understand what you want? If not, the model won't either.

Screenshot coming after our hands-on test
Screenshot of the same request before and after the method: coming after our hands-on test.

1. Context: who, for whom, why

Start with the situation. Who are you, who will read the result, and what is it for? Google's guide recommends including the information the model needs "instead of assuming that the model has all of the required information". Anthropic adds that explaining why you want something helps the model understand your goal and give a more targeted answer.

If you work from a document (an email, notes, a report), paste it. Anthropic recommends placing long documents near the top of the prompt, above your instructions.

2. Task: one clear verb

Say exactly what you want done: summarise, rewrite, compare, list, draft. OpenAI's help center asks for prompts that are "clear, specific" and that avoid ambiguity. If the task has several stages, number them, or split them into several messages: Google's guide suggests breaking complex prompts into simpler components, and Anthropic recommends numbered steps when order matters.

3. Constraints: what is in and out

Length, tone, level of detail, what to avoid. Google's guide shows how to tell the model "what to do and not to do", for example the length of a summary. OpenAI suggests describing tone with plain adjectives: formal, friendly, professional, serious. Anthropic notes that it works better to say what you want than only what you don't want.

4. Format: the shape of the answer

Ask for the shape you need: a table, a bulleted list, three short paragraphs, an email ready to send. Google lists these formats as examples. The format is also a quality check: an answer in a table forces clear categories.

Put together, the first four parts look like this:

I run a small bakery and I'm writing to regular customers who signed up for our newsletter (context). Write an email announcing our new Sunday opening hours (task). Friendly tone, under 120 words, no discount offer (constraints). Give me a subject line, then the email body (format).

5. Examples: show, don't only tell

An example is often worth a paragraph of instructions. Google's guide goes as far as recommending that you "always include few-shot examples" in your prompts. Anthropic calls examples "one of the most reliable ways" to steer format, tone and structure, and advises examples that are relevant to your case and varied enough that the model does not copy one detail by accident.

In practice, paste one or two past emails you liked, or a sample line of the table you want, and say: "follow this style".

6. Iterate: the first answer is a draft

OpenAI and Google both insist on it: prompting is iterative. OpenAI describes the loop plainly: start with a prompt, review the answer, then adjust the wording, add context or simplify. Google calls its guidelines "starting points". Rather than starting over, answer the assistant:

Good start. Make it shorter, remove the second paragraph, and replace the closing line with an invitation to reply to this email.

When a prompt works well for a task you repeat, save it in a document and reuse it.

Three limits to keep in mind

  • A good prompt does not guarantee a true answer. Check facts, figures and names, especially if you will publish or send the result.
  • Watch what you paste. Context helps, but avoid personal or confidential data you are not allowed to share.
  • Models change. The official guides are updated as models evolve. A prompt that worked last month may need a small adjustment.

Tools in this guide

Questions

Do I need a long prompt every time?

No. A quick question deserves a quick prompt. The method is useful when the first answer disappoints, or for a task you will repeat often.

Does the same prompt work in ChatGPT, Claude and Gemini?

The principles are very close in the three official guides: be clear, give context, show examples. Results still differ between models, so test your important prompts on the one you use.

Are there magic words that make AI answer better?

Not really. In its advice for Gemini 3 models, Google's guide even recommends avoiding unnecessary or overly persuasive language. Clear context and a precise request do more than any formula.

Sources

  1. Anthropic, Prompting best practices, Claude documentation
  2. OpenAI Help Center, Prompt engineering best practices for ChatGPT
  3. Google, Prompt design strategies, Gemini API documentation