Avatar

How to brief an AI now

TL;DR

  • The newer models follow your words closely. A thin brief now gets a thin answer.
  • A good brief has four parts: the task, why and who it is for, the rules with their exceptions, and what done looks like.
  • Give it all at once, at the start, and let the AI do the boiling down.

Most of us learned to talk to an AI the way we learned to talk to a search box. Short. A few keywords. Then try again.

That habit now costs you.

The newer models read what you write closely. They do what it says. Anthropic's own guide to its newest model gives one example: "The model treats an instruction that contains 'never' as absolute" (Prompting Claude Fable 5.2, as of November 2026).

So the brief is the job. What you leave out, the AI fills in with a safe guess. What you put in, it follows.

This guide is the short version of how I brief an AI today. The other six guides in this set each take one part of it further.

The four parts of a brief

A good brief answers four questions. I write them in this order.

1. What is the task? Say what you want made or done. One job, named plainly. "Write the opening page for a plumber's website" is a task. "Help with my website" is not.

2. Why, and who is it for? Say who will read it, use it or pay for it. Say what you are trying to get. The AI makes a hundred small choices on the way. Each one goes better when it knows the reader.

3. What are the rules, and their exceptions? Say what must be true. Say why. If a rule has an exception, write the exception down. A rule with no reason is followed too hard in the places it was never meant for. The next guide, Say why, and stop shouting, is about this part.

4. What does done look like? Say how you will check it. A word count, a file that must open, a question it must answer. If you can, show a finished example of the kind of thing you want. Show it, do not describe it is about that.

Give it all at once

The second habit to drop is feeding the job in slices. You ask for a draft. Then you add the reader. Then you add the rule you forgot. Each slice is a correction, and each correction undoes some of the work before it.

Give the whole job at the start instead. It can be long. That is fine.

I wrote about this in Give the AI the long letter. There is an old line about apologising for a long letter because there was no time to write a short one. The short letter was always the hard one, because someone had to boil the thinking down. With an AI, the boiling down is its job, and it does it well. My job is to give it everything I know.

I talk most of my briefs now rather than type them. I say far more that way, and what I say is richer.

Say what is yours to decide

The newer models are careful about acting on things that look like your call. Anthropic's guide says a vague rule about approval made the model hold back about half of its routine requests (Prompting Claude Fable 5.2, as of November 2026).

So say both halves. Name the routine steps it may take on its own. Name the few things it must ask you about first. That keeps it moving on the small things and stopping on the big ones.

Ask for the evidence

When the work comes back, you need to know what is true. So ask for the source of each number and each claim. Ask it to say which checks it ran, and what each one showed.

Do not ask it to write out its thinking. That kind of request can now be refused, and a source is more use to you anyway. Ask for the source, not the reasoning explains why.

A brief, start to finish

Here is a brief I would give for a small job.

Write the opening section of a website for a plumber in Bridgewater, Nova Scotia. The reader is a homeowner with a leak, often on a phone. They want to know two things fast: does this plumber do my job, and can I call now. Use only the services and the town I give you below. Do not promise a reply time, because one person answers the phone and cannot always keep it. Keep the headline to one idea, with no dashes. Done means: a headline, one short paragraph and a call button, and every fact on it is one I gave you.

That is four parts in eight sentences. Each rule carries its reason. And I can check "done" without asking anyone.

What this set covers

The other six guides each take one habit further:

Say why, and stop shouting is about rules.

Show it, do not describe it is about examples.

Ask for the source, not the reasoning is about checking the work.

Say exactly what you are saying yes to is about approvals.

Pick the model and the effort before you start is about cost.

Prune your instructions is for anyone who keeps a standing instruction file for an AI.

Every claim in this set links to its source, and each was checked in November 2026. These tools change often. I will check the set again when a new model comes out.

Related reading

Contact

Tell me what you’re working on.

Say what you are trying to fix and I will give you a straight read. If it is easier to talk, pick a time that suits you.

Book a time to talk

Prefer to write?

Shei Goldberg

I read every message myself, and I will get back to you.

We will text or call you on it.