The Anti-Magical Brief

A brief is the document that tells the AI — or the team building the AI — what good looks like. It is the single highest-leverage artefact in an AI deployment, and it is, almost without exception, written terribly.

Most briefs read like wishes. They describe the desired output in vague, aspirational terms — professional tone, on-brand, clear and helpful — and then trust the AI to fill in everything that wasn't specified. The AI cannot fill in what wasn't specified. Anything that wasn't specified gets filled in with the model's defaults, which are average, generic, and not yours.

I have started calling the alternative the anti-magical brief. It is built on a simple discipline: if the AI would have to magic something into existence to satisfy this brief, the brief isn't done.

Here is the brief I see most often. I'm not exaggerating it. It is the actual shape of what teams send to AI builds.

"Write professional, friendly customer support replies that match our brand voice. They should be clear, helpful, and not too long. Address the customer's question and offer a solution. Sound like one of our team."

Look at that brief from the model's perspective. Professional — by whose standard? Friendly — how friendly? Brand voice — defined where? Not too long — what's the cutoff? Sound like one of our team — what does your team sound like? The brief is one sentence of specification and four sentences of hand-waving, and the model fills the gaps with whatever average professional support reply it has seen most often. That's why your AI sounds like every other AI.

What's missing from the typical brief is everything that would let the AI produce something specifically yours. There are five categories of missing detail, and a serious brief addresses all of them.

Examples of good output. Three to five real examples of work you would sign off on, with annotations explaining what makes each one good. Not aspirational — actual.

Examples of bad output. Two or three examples of work that almost works but isn't acceptable, with annotations explaining what's wrong. The bad examples carry more information than the good ones, because they show the edge of the standard.

Acceptance criteria. The specific tests an output must pass. Length range. Required structure. Things that must never appear. Things that must always appear. The check the reviewer will run before approving.

Voice specification. Not friendly and professional — specific. Sentence length tendency. Vocabulary register. Things your team would say. Things they would never say. Punctuation conventions. Phrases that are signature. Phrases that are forbidden.

Failure modes. What should the system do when it encounters something it can't handle? Refuse? Escalate? Flag? Produce a template response? The brief that doesn't answer this question is the brief whose AI will produce confident nonsense in edge cases.

An anti-magical brief, when it's done, looks more like a style guide than a wish list. It contains pages, not paragraphs. It looks daunting at first.

And then a peculiar thing happens. Every brief after the first is shorter, because most of the work — the examples, the voice spec, the failure modes — was the work of the business, not the work of this specific brief. Once you've written it once, you reuse it. The next brief leans on the same artefacts. Over time, the briefs become small, sharp, and surrounded by a rich library of standards that every AI build references.

That library is, in my view, the most undervalued asset an SMB can build in the next two years. It is the institutional memory of what good looks like in your business, and once it exists, every new AI deployment starts from a strong foundation rather than from zero.

Writing an anti-magical brief feels slow. Founders, in particular, are allergic to it. Surely we can just point the AI at our existing material and let it figure us out? You cannot. The model has no way to distinguish your good work from your mediocre work, your historical work from your current standard, your aspirational tone from your accidental tone. Only you can do that, and the brief is the only artefact that does it.

The teams that take this seriously ship faster within three months. Not despite the slow start — because of it. The teams that try to skip the brief spend the next twelve months patching outputs, re-prompting, complaining about the model, switching vendors. None of it helps. The model was never the problem. The brief was.