How to brief an AI so the output is usable

The most common story about AI goes like this: people tried it and it wrote generic stuff. That's nearly always true, and it's nearly always down to the input.

A model doesn't know what it doesn't know. Give it no context and it produces the most general answer it can, because that carries the least risk. A generic answer isn't a fault. It reflects the information that's missing.

Five things every brief needs

Who it's for. Not "business owners", but "a managing director who already advertises and doesn't trust the report". The second one shapes word choice, length and examples.

What decision you want. What the reader should do once they've read it. Without that, the text describes something but takes nobody anywhere.

What the input is. If you have your own material, put it in. One real example is worth ten sentences of instructions. Whatever you leave out, the model will make up.

What not to write. People skip this one most often, and it's the most useful. Banned claims, banned phrases, banned formats. This is what keeps the voice yours.

What shape you want. Length, structure, how many points. Without it, the model falls back on its own default, which is the same one everyone else gets.

What makes the biggest difference

The negative instruction. If you only say what you want, the output drifts towards the average. Say what you won't accept, and it starts to sound like you.

One thing never to skip

Checking every factual claim. A model will write a convincing sentence around a number that doesn't exist. If a figure, a date, a regulation or a product spec shows up in the output, check it against the source. That's not distrust. It's the same check you'd run on a junior colleague's work.

A good brief isn't longer. It's more specific.