AI won’t save a bad brief

The fastest way to find out whether your brief is any good is to hand it to a machine.

For years, weak briefs got rescued without anyone noticing. A designer asked one sharp question at kickoff. The developer inferred what you probably meant. Somebody in a hallway said “wait, who is this actually for,” and the project quietly corrected course. The rescue was invisible, so the brief got the credit.

Amplifiers, not rescuers

AI doesn’t rescue. It complies. Feed it a vague input and it returns vague output at remarkable speed, formatted beautifully, delivered with total confidence. Garbage in, garbage out was always the rule; what changed is the throughput. A confused brief used to produce one wrong deliverable a week. Now it can produce forty before lunch, each one polished enough to survive a glance and hollow enough to fail an actual reading.

This is the part of the AI conversation nobody puts on a conference slide. The tools are amplifiers. They multiply whatever clarity or confusion you pour into them, and they do it without judgment, because judgment was never in the box. Hand them a sharp brief and they’re astonishing. Give them mush and you get mush at industrial scale, which is worse than ordinary mush because there’s so much of it to review, and reviewing bad options burns the exact hours you were hoping to save.

So the advantage has moved upstream. The unglamorous skill of the next decade isn’t prompting tricks or tool fluency. It’s the old, dull, decisive discipline of writing better inputs.

What a working brief answers

Here’s what we mean by better, because “good brief” has become one of those phrases everyone nods at and nobody defines. A brief that actually works answers a handful of questions so plainly that a stranger could act on them.

• Who is this for, and not as demographic soup but as one recognizable person with a problem.
• What single job is this thing supposed to do, because a deliverable with three jobs does zero of them.
• Which constraints are real, the budget and the deadline and the sacred cows.
• How will we know it’s done, described concretely enough that two people would agree when they saw it.
• What is deliberately out, because a brief that excludes nothing has decided nothing.

Notice that none of that requires software. It requires thinking, in advance, in writing, which is precisely why it’s rare.

A brief is thinking done ahead of time.

Most bad briefs aren’t badly written, they’re unwritten, a stack of assumptions wearing a template, and the author was hoping execution would surface the answers they didn’t want to sit with. Humans used to absorb that hope. The machine just hands it back to you, multiplied.

The loop got short

Our own ritual for this is unromantic. Before an ask goes to a tool or a teammate, someone who wasn’t in the room reads the brief and says back what we’re making. If their version doesn’t match ours, the brief loses, not the reader. Ten minutes of that has saved more hours this year than any model upgrade we’ve paid for.

We’ve started treating brief-writing the way we treat any craft: drafts, edits, reviews. When a project stalls now, the first place we look isn’t the output. It’s the input. Nine times out of ten the confusion in the deliverable is a faithful, high-resolution copy of the confusion in the ask. That was true before AI. The difference is that the loop used to take three weeks and now takes three minutes, so you meet your own vagueness almost immediately, wearing nicer typography.

There’s a strange gift in that, if you can stomach it. Fast, literal execution turns out to be a diagnostic. It shows you exactly what you asked for, which is not always what you wanted, and the gap between those two things is a precise map of what you haven’t figured out yet. Teams that read the map get sharper. The ones that blame the tool get faster at being wrong.

The industry will keep arguing about which model is smartest. Fine. We’d rather compete on the question the models can’t answer, the one that sits blinking at the top of every empty document before any tool gets involved.

What, exactly, are we trying to do here?

Answer that well and almost anything downstream can help you. Get it wrong and nothing can, though everything will now fail more efficiently than ever.