How to ask AI a better question

Most people’s disappointing experiences with AI come down to the same thing: a short, vague question producing a short, vague answer.

These tools respond to context. Give them almost none and they produce something generic. Here is what actually changes the result.

1. Say who it is for and what it is for

Compare:

Write something about our new opening hours.

with:

We are a small hardware shop in Warners Bay. Write a short notice for our front window telling customers we now open at 7am on Saturdays from 1 September. Friendly, not corporate, about 40 words.

The second produces something usable. The first produces filler.

The pattern is: who you are, what you need, who will read it, how long, what tone.

2. Give it the raw material

The single biggest improvement available. These tools are far better at working with information you provide than at recalling information themselves.

Instead of asking what a contract clause means, paste the clause. Instead of asking how to reply to a complaint, paste the complaint (with names removed). Instead of asking what to put in a report, paste your notes.

This also dramatically reduces invention, because it is working from something real rather than reaching for what such a document usually contains.

3. Tell it what “good” looks like

If you have an example, use it.

Here is a letter I sent last year that worked well. Write a new one in the same style about a different issue.

If you do not have an example, describe the qualities you want: plain English, no jargon, short sentences, firm but polite, suitable for someone with no technical background.

4. Ask for the reasoning, not just the answer

For anything where being right matters, ask it to work through the problem rather than state a conclusion.

Explain your reasoning step by step before giving the answer.

This genuinely improves accuracy, and it lets you see where it has gone wrong instead of having to take the conclusion on trust.

Related and useful:

Which parts of this are you least confident about?

Imperfect, but it flags the shakier claims more often than not, which tells you where to check.

5. Keep going

The first answer is a starting point, not a verdict. The conversation is where the value is.

• “Too formal. Try again, more like how you would say it out loud.”
• “Shorter. Half the length.”
• “That third paragraph is wrong. It was 2024, not 2022. Redo it.”
• “Give me three different versions and I will pick.”

That last one is worth doing often. Asking for options rather than an answer produces better results, because you get to exercise judgement rather than accept the first thing offered.

Things that do not help

Being polite. It makes no difference to the output. Say please if you like, but it is for your benefit, not its.

Elaborate “prompt formulas”. There is a small industry selling magic phrasings. The genuine improvements are the five above: context, material, examples, reasoning, iteration. Everything else is decoration.

Insisting it is wrong when it is right. These systems fold under pressure and will often agree with you rather than hold a correct position. That makes them poor judges of who is right, so do not use agreement as confirmation.

Very long questions with everything at once. Break a complicated task into steps and do them one at a time.

A template that works for most things

For anything of substance:

Context:  who I am and what the situation is
Task:     exactly what I want produced
Audience: who will read or use it
Format:   length, structure, tone
Material: the actual documents, notes or data

You do not need to label it like that. Just make sure those five things are somewhere in what you type.

An example worth stealing

Instead of “help me with my resume”:

I am 34, I have been a warehouse supervisor for eight years, and I want to move into IT support. I have just finished a Certificate III and I have built a home lab running a hypervisor and a few virtual machines. Here is my current resume [pasted]. Rewrite the summary and the experience section so the transferable skills are obvious to an IT hiring manager, without pretending I have professional IT experience I do not have.

That produces something genuinely useful, because it contains the real situation, the real constraint, and an explicit instruction not to overstate.

Practising

The fastest way to get good at this is to take a task you would have to do anyway, ask badly, then ask again properly, and compare. The difference is usually startling, and once seen it is hard to go back to one-line questions.

If you would rather work through this on your own real tasks with somebody beside you, that is exactly what the AI lessons are for, and the Thursday evening class covers it as a group.