How to write a good prompt for AI: the method

In brief

A good prompt says four things: who you are writing for, what you want to get, the useful context, and the expected format. Write it the way you would explain the work to a new colleague who knows neither your company nor your habits. Then correct the answer one point at a time, instead of starting over.

The four parts of a good prompt

The guides published by Anthropic, Google, and OpenAI say the same thing in different words: an AI does not know what you do not tell it. Anthropic advises thinking of it as a brilliant new employee who knows nothing about how you work. Google boils an effective prompt down to four parts, with no obligation to use them all every time.

Part	The question to ask yourself	Example
The role and the reader	Who is writing, and for whom?	I am the customer service manager. The message is for an unhappy customer.
The task	What exactly do I want to get?	Write a reply that acknowledges the delay and offers a solution.
The context	What does it need to know to do this well?	The order is ten days late. We can offer free shipping, not a refund.
The format	What should the result look like?	120 words at most, a polite and formal tone, a single question at the end.

In its April 2024 guide, Google notes that the most effective prompts average 21 words, while most people write fewer than nine. So the usual problem is not saying too much, but not saying enough.

Before and after: three examples

A follow-up email

Before: “Write a follow-up email for a quote.”

After: “I am the sales manager at a carpentry company. Write a follow-up email to a customer who received our quote ten days ago and has not replied. Goal: get a phone call scheduled this week. Friendly and direct tone, formal, 120 words at most, with a subject line. Do not make up any discount or deadline.”

Meeting notes

Before: “Summarize this meeting.”

After: “Here are my rough meeting notes. Turn them into a summary for the people who were absent: first the decisions made, then a table of action items with an owner and a date. A proposal must not become a decision. When a piece of information is missing, say so instead of guessing.”

Analyzing a spreadsheet

Before: “Analyze this file.”

After: “Here is the export of our sales for the year. Give me the three trends to remember, in plain sentences, for an executive committee. Flag the rows that look abnormal and say which columns you are relying on. End with the two figures I should check myself.”

In all three cases, the second version takes no special talent. It answers the questions a colleague would have asked you before getting started.

Five habits that change the answer

Say why. Anthropic recommends explaining the reason for a rule rather than just imposing it. “The text will be read aloud in a meeting” helps more than “keep it short.”
Show an example. Paste an email you have already written and ask for the same tone. To pin down a format, Anthropic’s documentation suggests three to five examples.
Ask it to question you. “If essential information is missing, ask me up to three questions before you write.” You avoid an answer built on guesses.
Forbid making things up. “Do not add any figure that is not in the document.” This does not excuse you from checking, but it cuts down on bad surprises.
Write in full sentences. Google advises natural, precise language without jargon: write as you would to a person, not as you would in a search engine.

A simple test, suggested by Anthropic: show your prompt to a colleague who does not know the file. If they do not understand what is expected of them, the AI will not either.

Correct rather than start over

The first answer is a draft. OpenAI and Google describe the same approach: read, say what is wrong, ask again. It is a conversation, not an order placed once and for all.

Name one precise flaw at a time: “too long,” “too casual,” “the second paragraph repeats the first.”
For tone, an adjective is often enough: more formal, warmer, more direct.
Keep what is good: “keep the introduction, rewrite only the conclusion.”
A new topic, a new conversation: the thread stays readable, and a very long conversation uses up your usage limit faster.

And always review the result before you use it. A careful prompt improves the form; it guarantees neither a figure, nor a date, nor a quote.

Keep what works

A prompt that gave you a good result is worth keeping. Replace what changes from one time to the next with brackets, for example [customer], [date], [amount], and store it where you will find it again.

In ChatGPT, a project keeps your files and your prompts for a topic that lasts.
In Gemini, a Gem is an assistant set up once for a task that comes back. Google plans to convert Gems into skills starting in November 2026.
In Mistral Vibe, a “Skill” saves a way of doing things: a memo template, a checklist, a tone.
In Claude, projects organize your files and documents, with the Pro plan.

A prompt should not be a suitcase for data. Before you paste a document into one, read what not to share with AI at work.

What Nova brings

Nova is subscription software. Each demo comes with its prompt ready, where all you have to do is fill in the brackets, and “My methods” keeps the ones that work for you. You learn the method by practicing it on your tasks, not by reading a course. Try the interactive demo.

Frequently asked questions

Do prompts need to be long?

Not long: complete. According to Google’s April 2024 guide, the most effective prompts average 21 words, and most people write fewer than nine. Add the reader, the context, and the format, then stop.

Should you give the AI a role?

It helps, without being required. Anthropic says a single sentence of role already makes a difference in tone and behavior. Google counts it among its four parts, and specifies that you do not need to use them all.

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

Yes, for the most part. The three vendors’ guides rest on the same principles: be clear, give the context, specify the format, then refine. A prompt written well for one generally gives a good result in the others.

What should you do when the answer misses the point?

Do not start from scratch. Say what is wrong, one flaw at a time, and give an example of what you expect. If the thread gets confusing, open a new conversation with a more precise prompt.

Can you paste a company document into a prompt?

Only if you have the right to share it and the tool is approved by your company. Otherwise, remove the names, contact details, and amounts. If in doubt, ask your data protection officer.

Sources

Prompting best practices, Anthropic, Claude documentation
Prompting guide 101 (April 2024 edition), Google
Prompt engineering best practices for ChatGPT, OpenAI, help center

How to write a good prompt: the method, with examples