A brilliant marketer I know told me she'd 'tried AI and it wasn't good.' Her prompt had been four words: 'write a launch email.' What came back opened with 'We're thrilled to announce' and closed with 'Don't miss out' — competent, weightless, and about a product it plainly knew nothing about. We spent ninety seconds rewriting the request the way she would brief a new hire: who to be, what the job was, who it was for, what 'done' looked like, and one line inviting it to ask her questions first. It asked two. Her answers were the parts of the pitch she had never written down anywhere. The draft that followed opened on the thing freelancers actually hate about tracking expenses. Nothing about the model changed. The brief did.
The five-line brief
Four of those five lines are the ones everybody works out eventually on their own. The fifth is the one almost nobody writes, and it does the most work. Asking a model what it needs to know before it starts costs you a single exchange and buys you the gap analysis you were never going to run on yourself. Half the time its questions are the brief: you find out what you actually wanted by answering them, which is exactly what happens when you hand a job to a competent person and they push back before touching it.
Write the brief, line by line
- Open by telling it who to be. That one line sets the vocabulary, the judgment and the standards for everything that follows.
- One request, one outcome. If you need two things, that's two briefs.
- Audience, hook, tone. This is where generic becomes specific, and it's the only part of the brief nobody else could have written.
- Length, format, and what to avoid. Constraints don't limit the model; they aim it.
- 'Ask me up to three questions before you start if anything's unclear.' This is the line almost nobody uses and everybody should.
- Steer in fragments — 'warmer', 'half as long', 'the second paragraph is the real opening, start there.'
The other half of the skill is refusing to treat the first draft as a verdict on the tool. A draft is a probe. React to it in fragments and each reaction costs about five seconds and lands closer than the one before, because you're steering with the destination in view instead of guessing at it from the driveway. Three rounds of that beats an hour spent engineering a perfect opening prompt, and it beats it on wall-clock time as well as on the result. The people who find these tools disappointing are almost always the people who read draft one, decided, and closed the tab.
The four ways a brief fails
Two honest limits before you go and use this. It's a working method, not a tested finding: it comes out of daily use across the models this publication covers, and the effect size is not established — we link no study here because there isn't one we can point you at that tested this specific skeleton. And no brief rescues a job that is really two jobs; if your request has an 'and' in the middle, split it or accept two mediocre halves. What the method does do, reliably, is remove the most common reason people conclude that AI 'isn't good'. They asked for something generic. Generic is precisely what these systems are best at producing on demand.
- The highest-leverage AI skill is briefing like a manager, not prompting tricks.
- Give a role, state one job, add the two or three facts only you know.
- Set constraints — length, format, what to avoid. Constraints aim the model.
- End by asking the model what it needs to know before it drafts anything.
- This is a working method, not a tested finding — we link no study, because there isn't one.