How to write an effective ChatGPT prompt: method and examples

A restaurant owner types: "Write a Facebook post for my restaurant." ChatGPT replies with "authentic" cooking, a "warm" atmosphere, and three hashtags. Nothing wrong with it, but the text could be used for any restaurant in the country, and she's going to rewrite it entirely.

Here's what she could have sent instead:

I run a 30-seat crêperie in Vannes, popular mostly with regulars from the neighborhood.
I want to fill the Tuesday evening service, the slowest of the week.

Write a Facebook post announcing our Tuesday deal: a complete galette, a homemade salted caramel crêpe, and a bowl of cider, for €[PRICE].

Constraints:
- 80 words maximum, simple and warm tone, as if I were talking to my customers in the dining room
- No superlatives and no string of exclamation points
- End with the reservation number: [PHONE]
- Don't invent any detail about the restaurant that I haven't given you

Give me 3 versions with three different hooks.

There's nothing clever about this prompt. It says what she's after (filling a Tuesday evening), supplies the facts ChatGPT can't guess, and sets limits. Everything else in this guide follows from that logic.

Describe the result you want before describing the method

In its advice for ChatGPT, OpenAI recommends starting with the desired outcome rather than a detailed list of steps, then setting clear limits. That's a change from the advice you read everywhere, which consists of stacking up roles and instructions.

A role is still useful when it really changes the point of view, and it costs one line. But "You are a marketing expert with 20 years of experience" adds nothing if you don't say what your text needs to achieve. Try it with and without: if the answer doesn't change, drop it.

To write that outcome, ask yourself one question: what would make me say "that's it, I can use this"? For an accountant preparing a reminder email, "a polite email" doesn't say much. "A message the client reads to the end and that gets them to send their supporting documents before the 15th" changes everything: ChatGPT puts the deadline and the requested action up front instead of burying them in pleasantries.

A simple framework comes down to four blocks:

GOAL: [what the result should let you do, and for whom]

CONTEXT: [the facts ChatGPT can't guess: situation, audience, real constraints]

LIMITS: [what it should avoid, the sources to use, what it should do if it's missing a piece of information]

FORMAT: [length, structure, tone, an example if you have one]

You don't need these headings every time. For a simple question, one sentence is enough. The framework mostly serves as a checklist when the result disappoints you: first see whether one of the four blocks is empty. The article on the 6 elements of a good prompt goes deeper into each of them.

Set the limits that stop ChatGPT from making things up

ChatGPT's costliest flaw is the way it fills gaps with something plausible. If you ask for meeting minutes from incomplete notes, it may add a decision that was never made, with the same confidence as everything else.

You can't eliminate this risk. You can reduce it by giving an instruction for when information is missing. OpenAI also suggests limiting the model to the sources you provide and asking it to flag what's missing instead of guessing. In practice, for an association treasurer preparing her general meeting:

Here are the raw notes from our last board meeting:

[PASTE THE NOTES]

Write a one-page summary for the members.

Rules:
- Use only what appears in these notes. Don't add any decision, date, amount, or name.
- When a piece of information seems necessary but is missing, write [TO BE COMPLETED: what is missing] at that spot.
- Clearly separate decisions made, topics discussed without a decision, and actions to take, with the person responsible if one is given.

After the summary, list the passages in my notes that you found ambiguous and explain how you interpreted them.

The last instruction is the one that saves the most time. Instead of rereading the whole text with the same level of attention, you go straight to the fragile spots. You still have to reread everything if the document commits the association, but you know where to look first. Other techniques exist, found in the article on reducing hallucinations.

Fix things in small touches, and know when to start over

The first result is a draft. Rather than replying "it's not bad but it's too long and not direct enough," name the flaw and what you want instead.

A conversation that drags on ends up working against you, though. After ten back-and-forths, ChatGPT keeps in context attempts you've rejected, and they keep rubbing off on its answers. When you feel that every fix breaks another one, ask it to write a prompt that pulls together everything that's been established, then paste it into a new conversation. That often works better than an eleventh adjustment. The full method is in the guide to improving a prompt through iteration.

Use the right settings for what needs to last

Plenty of people re-paste the same instructions into every conversation even though ChatGPT offers several places to store them. Their scope isn't the same, and the names and availability depend on your plan: at the time of writing, check in the tool.

Custom instructions should stay short and general. Put in them what's true for almost all your requests, for example: "I'm a self-employed electrician in Brittany. Answer in French, with no introduction, and tell me when you're not sure." What concerns only one job site or one client goes in a project.

Custom GPTs are on their way out. OpenAI has announced their retirement in favor of plugins, in which a GPT's instructions become a skill and its knowledge files become reference files. The date and conditions vary by account type, so check the migration FAQ in OpenAI's help center. If you were about to create a GPT, start with a project or a plugin instead. If you use one built by someone else, you can't migrate it yourself: look at the notice displayed on the GPT and ask its author what they plan to do. The article on system instructions and projects explains how to write them.

Frame a deep research request before launching it

Deep research deserves a prompt of its own. It will comb through many sources and hand you a cited report. If your question is fuzzy, you lose several minutes to get a generic report, and the number of searches may be limited depending on your plan.

Weak: "Do a market study on heat pumps."

Better, for a heating technician who's hesitating about taking the leap:

I'm a self-employed heating technician in [DÉPARTEMENT], with no employees. I'm considering adding air-to-water heat pump installation to my business within the next 12 months.

Prepare a report that helps me decide. It must answer these questions:
1. What qualifications and certifications are required for my customers to access the current public subsidies?
2. What are the training and equipment costs to get started on my own?
3. What problems are most often reported by customers and installers?
4. What signals indicate that demand is rising or falling in my region?

Instructions:
- Favor official French sources and professional bodies. Give the date of each piece of information.
- Separate what is established by an official source from what comes from opinions or commercial articles.
- If two sources contradict each other, show both instead of settling it.
- End with what you were unable to verify.

Start by telling me how you plan to go about it and wait for my approval.

That last sentence applies to any long job: OpenAI advises asking for a plan when the way of doing it matters. Fixing a plan takes thirty seconds. Fixing a full report means rerunning everything.

Once you've received the report, open the links that support the points you're going to base your decision on. A citation shows where a piece of information was taken from; it doesn't prove the information was understood correctly or that it's still valid. For public subsidies and professional rules, check the relevant official website or your professional body.

Test your prompt before you rely on it

A prompt that worked once isn't yet a good prompt. It may simply have hit an easy case. If you plan to reuse it, try it on three real situations: a normal case, a case where information is missing, and a tricky case. For the crêperie, that would be a standard deal, an announcement where you don't give the price, and an exceptional closure.

Pay special attention to what it does on the incomplete case. If it invents the price, add the rule that stops it. Then save the version that works, with [BRACKETS] where the values change, in a document of your own: you'll be able to reuse it without starting from scratch.

The same prompt can also give a different result from one time to the next, and a model change in ChatGPT can alter its behavior. So rerun your important prompts from time to time, especially after an update. For repeated work, imposing a precise format limits the variation.

To see ready-made prompts and their actual output before deciding to buy one, the YouPrompt catalog shows an excerpt and a real result for each listing.

Now open your ChatGPT history and find the last answer you had to rewrite. Take the original prompt, spot which of the four blocks was empty, fill it in, and rerun it in a new conversation. Compare the two answers side by side.