Prompt: how do you ask for exactly what you want?

The short version

Asked ChatGPT for something and got back a generic reply you couldn't use? Before you judge the tool, look at what you wrote. Very often the problem isn't the answer; it's that the request itself was missing something, without you noticing.

In this lesson we give ChatGPT the same task twice: once with a short, incomplete request, which gets us back a template full of blanks we have to fill in ourselves, and once with a request that includes what it needs, which gets us a reply much closer to what we want. Same tool, same task; what changed is the Prompt.

Key terms

• Prompt: the text or instructions you send to the model.

The explanation

The slip you hand in at the counter

Go back to the government office counter from the first lesson: behind the counter sits a clerk, the Model, and you write your request for them on a slip of paper.

You wrote on the slip: "Write me a leave request", and handed it in. The clerk looks at the slip: leave when? Addressed to whom? For how many days? How formal should it be? None of that is written down.

They now have a few options: ask you, fill in the details on their own, or hand you a blank form to complete yourself. Had the important details been on the slip from the start, they could have got to work straight away. Same clerk, same counter; the difference is the slip.

What is a Prompt?

That slip is the Prompt: everything you write to the model to ask it for something. It doesn't have to be a question, and it doesn't have to be long. You might ask it to write, to summarize a text, or to fix some code; all of that is a Prompt.

And the better Prompt isn't necessarily the longer one. This short request, for example, is perfectly enough if you paste the text you want summarized under it:

Summarize this text in three bullet points.

What matters isn't length; it's saying the things that make a difference to the result.

Four questions for when the result isn't what you wanted

If you get a result that doesn't suit you, check your request against these four questions:

• The task: what do you want it to do?
• The information: what does it need to know to get the task done?
• The constraints: what must it stick to or avoid, such as the tone?
• The shape of the result: how do you want the answer to look: its length, its language, its structure?

Example: a leave-request email

We picked writing as the example because the difference shows up quickly. Look at this request:

Write me a leave email

The task is there: write an email. But the information is missing: leave when? To whom? The constraints are missing: what tone? And so is the shape of the result: how long? In what language?

When something important is missing, the model will either ask you, fill it in on its own, or leave you a blank to fill. Asking isn't a mistake; it's better than guessing.

In the video we sent this request in a new chat, and ChatGPT returned a ready-made template with three blanks in brackets: [Start date], [End date] and [Your name]. In other words, part of the work came back to us.

Then, in a second new chat, we sent the same task along with what had been missing:

Write me a short email in Modern Standard Arabic, exactly 4 lines, from an employee to their manager, requesting next Thursday off for a family matter, in a respectful, understated tone, and include a subject line

Here is what each part of it covers:

This time the reply came back with no blanks and stuck to the constraints, so it needed far less editing. It wasn't entirely ready to send without any touch-up (it didn't include the sender's name, for instance), but it was much closer to what we asked for.

Not just writing

The same idea applies to anything you ask of ChatGPT or any other tool:

• Images: "Make me a picture of a desk" is an understandable request. But if you want an engineer's desk, at night, with soft lighting, in landscape format, say so.
• Analysis: "Analyse these numbers" may get you generalities. "Pull out the three biggest drops in sales and show them in a table" is a far clearer request.
• Coding: "Make me a website" is completely different from "Make a single page for a restaurant that works well on a phone, with the menu and contact details, and no sign-in".

The principle is the same: if a detail makes a difference and you didn't mention it, the model won't know it's in your head. It will either ask you, guess, or give you something generic.

Try it yourself

Write me a leave email

Write me a short email, exactly 4 lines, from an employee to their manager, requesting next Thursday off for a family matter, in a respectful, understated tone, and include a subject line

In Claude

The same steps work in Claude, but it requires signing in. So use a clean session: turn off memory and personalization, don't attach any files, don't pick a project, then send each request in a new chat. Claude may also ask you about the missing details in the short request instead of writing the message straight away.

Why this matters to you

When you start building real things with coding tools, a large part of your work will be explaining exactly what you want, then reviewing what was done. The tool won't know that the page has to work on a phone, or that it doesn't need a sign-in, unless you tell it. Every missing detail in your request turns into a guess in the result, and then into time spent fixing it.

Related lessons

• What is AI?: what Model means, and the government office counter.
• How does AI think?: what a Token is, and how the model reads your request and writes its reply.