Asking well

Ever asked the same thing twice and got two completely different answers?

Let's try it.

You asked exactly the same thing

Add like a teacher and the words get simpler.

Add like a scientist and technical terms appear.

The question never changed. You put one phrase in front of it.

What you tell an AI to do, or ask it, is called a prompt.

So why does it change?

Remember 4.2. A language model does one job: guess the next piece.

A prompt is text that goes in front.

"Like a primary school teacher, explain why volcanoes erupt." What kind of words are likely to come after that sentence?

Simple ones. In the writing it learned from, that spot was nearly always followed by simple words.

The AI didn't pick a tone for you. You changed what came before, so what came after changed.

A good request has three things in it

Count what you changed in the lab and there are three.

• Who it's talking to — like a teacher, like a scientist

• How much of it — three lines, one line

• What shape — a paragraph, a list

They're the same three you'd use on a person. "Keep it short." "Explain it simply."

So what can you actually give it to do?

You've seen it handle words. Let's try something else.

Reading it told you nothing

You asked for three steps and got four.

Nothing looked odd while you were reading. Not until you counted the repeats and found 4.

One number was wrong. And reading it never showed you.

The star was one angle. The rainbow was one colour.

You have to run it

The point here isn't code.

It's that what an AI makes looks right as long as it reads well. Exactly like the book that didn't exist.

With one difference. Code you can run.

Checking the book means digging through a library. Checking the code takes one button.

But the square was fine

One of the four had nothing wrong with it.

Same lesson again. It isn't wrong every time. That's precisely why you check.

How to decide what to hand over

This guide offers one test. Can I check the result myself?

Handing over something you can't check, and then believing the answer — that's the dangerous spot.

Try it yourself

Write down your best request as a single line.

Next time you use an AI, start with that line.

That's the talking AI

Cutting text into pieces, guessing the next piece, choosing where to look — you've seen all of it.

Next comes AI that makes things. Pictures and music.