What a prompt is and how to write a good one, with before-and-after examples

A prompt is the message you write to an AI such as ChatGPT, Gemini or Claude to ask it for something: a question, an instruction or a full brief with every detail. Writing a good prompt means giving the AI what you’d give a new hire: the context, the exact task, the format you expect and, if you can, an example. With those, the answer stops being generic.

The official guides from Anthropic, OpenAI and Google agree on those four pieces, and on one more thing: check what comes back, because AI can make things up. Below you’ll find before-and-after examples (an email to a customer, a restaurant menu, a homework assignment), what doesn’t help, such as threats or tips, and how much longer a complete prompt really is. We measured it.

What is a prompt?

A prompt is the input an AI model receives: everything you type, say or attach before it answers, on your phone or your computer. It can be a five-word question or a full page with data, examples and rules. Google’s guide for Gemini in Workspace describes it as a conversation starter with your AI assistant.

How does an AI understand what you ask?

Picture a brilliant employee on their first day: they know a lot, but not your customers, your tone or what the job is for. That’s the comparison in Anthropic’s prompting guide (Anthropic makes Claude), which suggests a simple test: show your prompt to a colleague with no context. If they’d be confused, the AI will be too.

The AI doesn’t guess what you leave out: it fills the gaps with the most likely option, which is usually the most generic one. OpenAI says the same in its help center: models often do best when you ask as you would ask another person, with clear details. Everything you write is read in small pieces called tokens; we explain them in what tokens are in AI.

What are the parts of a good prompt?

A good prompt has four parts: context (who you are, who it’s for and why), task (what to do, with a clear verb), format (length, tone, structure) and examples (one or more models of what you want). You don’t always need all four, but every missing one is a gap the AI fills its own way.

• Context. Who will read the result, where it will be used and what «done well» looks like. Anthropic also asks you to explain why: instead of «NEVER use ellipses», its example says the text will be read aloud by a voice engine that can’t pronounce them.
• Task. One concrete action: «draft», «summarize in five points», «compare». If there are several steps in order, number them.
• Format. Length, tone, language and shape: one sentence, a table, three options. OpenAI’s best practices ask you to be specific about the outcome, length, format and style.
• Examples. One model of what you expect beats three adjectives. Google goes as far as recommending them in every prompt for Gemini.
• And what almost nobody writes: limits. What it shouldn’t do and what to do when a detail is missing, for example «if you don’t know, say so».

Prompt examples: before and after

An email to a customer

Before: «Write an email to a customer saying their order is delayed.» The AI doesn’t know how late it is, why, what you’re offering or how you talk to customers, so it invents a generic apology, maybe with a discount you never planned to give.

After: «I run an online store that sells solid-wood furniture. Write an email to Ms. Patel, a customer for two years: her dining table will arrive on November 6 instead of October 30 because the finish needs more drying time. Offer free shipping as compensation, nothing else. Warm and direct tone, 120 words max, with a subject line. If any detail is missing, leave it in brackets.»

Everything changes: the AI now has the date, the reason, what to offer and what not to, the tone and the length. The brackets stop it from filling in what you didn’t give it with made-up details.

A dish description for a menu

Before: «Write a description of breakfast tacos for my restaurant's menu.» You’ll get generic tacos, with ingredients you may not use and the tone of a travel brochure.

After: «I run a family café in Austin, Texas, and I'm redoing the printed menu. Write the description for “House breakfast tacos”. Ingredients: scrambled eggs, chorizo, potatoes, cheddar and flour tortillas; served with salsa verde. Price: $8.50. Format: one sentence, 25 words max, warm tone, no exclamation marks, price at the end. Use only the ingredients I give you. Example of my menu's style: “Buttermilk pancakes with maple syrup and a pat of butter, just like Sunday mornings. $7.00”»

Here the example does the hardest part: it shows your menu’s voice better than any adjective could. And «use only the ingredients I give you» keeps out the avocado or the bacon you don’t serve.

A homework assignment

Before: «Do my fractions homework.» You get answers that may not follow the method taught in class, and the student learns nothing.

After: «My son is in fifth grade and has to add fractions with unlike denominators. Explain the least common multiple method with a pizza example, in words a 10-year-old understands. Then give him three exercises without the answers and, separately at the end, the answers so I can check his work.»

How much longer is a good prompt? We measured it

On October 25, 2026, we counted the words and tokens of the two menu prompts above with gpt-tokenizer 4.1.0, an open-source library that reproduces OpenAI’s o200k_base tokenizer, the one used by GPT-4o.

The complete prompt is almost eight times longer and takes a couple of minutes to write. In return, it can save you the three or four rounds of «shorter», «no exclamation marks», «we don’t serve that». Claude and Gemini split text with their own tokenizers, so their counts will differ a little.

How do you get the format you want?

Say what you want in positive terms. Both OpenAI and Anthropic advise telling the model what to do instead of what not to do: «write in flowing paragraphs» works better than «don’t use bullet points».

Show before you explain. One example conveys length, tone and structure at once. Anthropic suggests 3 to 5 varied examples for repeated tasks, set apart with tags such as <example>; Google warns that with too many, the model starts copying them instead of picking up the pattern.

Do «magic prompts», threats or tips work?

Not reliably. A Wharton School study tested threatening an AI and offering it a tip, and in general neither had a significant effect on its results. No secret formula replaces what actually changes the answer: context, a clear task, the format and examples.

The report, published in August 2025 by Lennart Meincke, Ethan Mollick and colleagues, was sparked by Google co-founder Sergey Brin saying in May 2025 that models do better if you threaten them. They measured it on two sets of hard questions, GPQA and MMLU-Pro. Some variations changed the result of a single question a lot, but there was no way to predict whether they would help or hurt.

Shouting doesn’t help either. Anthropic’s guide says its recent models follow instructions more closely and recommends softening orders like «CRITICAL: You MUST…» into a plain «Use this when…», because alarmist wording can make the AI apply the rule too often.

• «Copy this perfect prompt». Templates help you not forget a part, but the value lies in what only you know: your customer, your dish, your date.
• «Tell it it’s a world-class expert». A role does help set the tone and focus, according to Anthropic, but «you’re the best lawyer in the world» doesn’t give it knowledge it lacks.

How do you check what the AI tells you?

Treat the answer like a new colleague’s draft: check figures, names, dates and quotes against the original source before using it. The companies themselves ask for this. Google’s guide reminds you that the final result is yours, and Anthropic warns that its techniques reduce made-up answers but don’t eliminate them.

Some checks can go right into the prompt, according to Anthropic’s guide to reducing hallucinations:

1. Let it say it doesn’t know: «if you don’t have the information, say so». Anthropic says this can drastically reduce false information.
2. Tie it to your document: paste the text and ask it to quote the exact sentences it relies on.
3. Make it review its own answer: it should find a supporting quote for each claim and drop any claim without one.
4. Check outside the AI whatever matters: prices, laws, doses, anything about health or money.

Images work the same way: an AI can produce a believable photo of something that never happened. What to look for is in how to tell if an image is AI-generated.

What changes for you and your business?

If you use AI to write emails, posts or product descriptions, a good prompt can be reused. In practice:

• Save the prompts that worked in a document, with blanks for what changes: [customer], [date], [product].
• Don’t paste what you wouldn’t share: passwords, bank details or your customers’ private information.
• Ask for the language and the region: «in American English, friendly but professional» avoids words your customers don’t use.
• Not everything needs a chat. To translate a phrase and hear how it’s pronounced, our free translator is quicker.

What to remember

A prompt is the brief you give the AI, and it works better the better you explain it: context, task, format and an example. Threats and magic formulas don’t replace that, and the answer always gets checked.

For jobs that don’t need a chat (scanning a document, making a QR code, removing a photo’s background), there are our tools, several of them free and with no sign-up.

Frequently asked questions