Why every first draft is generic
The model fills in missing information with the average. If you do not say who it is for, it will write for everybody. If you do not say what your intent is, it will inform. If you do not say what it should leave out, it will write in everything that appears in every similar text.
That is why the brief is not admin, it is most of the work. Writing a good brief takes fifteen minutes; rewriting the text that came out of a bad brief takes two hours. The difference does not come from the model, it comes from how many decisions you made beforehand.
The seven fields
- Who it is for. Not "SME managers", but "the owner of a five-person accounting firm who has just heard about AI for the first time and is afraid of being left behind".
- What they should know at the end. One single sentence. If you cannot put it in one sentence, the text will not hold together either.
- What they should do next. The action you are hoping for. This sets the ending and the tone.
- What they already know. The assumed prior knowledge. Without it the text will be either patronising or incomprehensible.
- What voice. Not with adjectives ("friendly, professional"), but with samples: five paragraphs from your own earlier material.
- What the specific is. At least one number, one case or one example that you supply. This is the thing the model cannot invent, and it is what makes the text yours.
- What it should not write. The banned list. Phrases, platitudes and topics you do not want to see. This is the most frequently skipped field and one of the most effective.
What there is no measurement for — and let us say so
We have to be honest here: we found no study that measured, field by field, how much each brief element improves things. There is no ablation research on what specifying the audience is worth compared with specifying the tone.
So the seven fields above are an experience-based list, not a measured ranking. That is worth saying out loud, because the claims going round on this subject — the "a good brief makes the quality 3.4 times better" type — have no such measurement behind them either. What practice does show consistently: the model makes up for missing information with the average, and the more decisions you take in advance, the less you have to fix afterwards.
What there is measurement for
On one thing there is good, large-sample data: on tone. According to a study across six models and forty thousand texts, five of your own samples are enough to carry the style over — and ten make it no better. And, against common sense: if you pick the samples by topic (that is, you supply earlier texts of yours about the same thing), that makes the result worse.
The same study also showed where the ceiling is: for news-like, factual text the style match is 86 to 93 percent, but for a personal-voice blog it is 38 to 44. So the tone field is worth less the more personal the text is — and your own voice is still something you have to put in yourself. The Train the AI on Your Past Writing post goes through this in detail.
The template
AUDIENCE: [one specific person, in a specific situation]
GOAL: by the end they know that _______________
ACTION: afterwards they go and _______________
KNOWS: already knows _______; does NOT know _______
TONE: see the 5 paragraphs below from our earlier material
(picked at random, NOT by topic)
SPECIFIC: [number / case / example that I supply]
BANNED: do not use: [phrases]; do not write about: [topics];
do not open with [introductory platitudes]
LENGTH: _____ words, _____ sections
What to leave out of it
- The keyword list. Put it in the brief and the text will be built around it, and it will show. The keyword comes up at editing, not at writing.
- The "make it SEO-friendly" instruction. It means nothing, and the model will fill it with platitudes.
- A competitor's text as a sample. The draft will end up resembling it — which is the opposite of what you want.
- A pile of adjectives for the tone. "Friendly but professional, direct but expert" — the model makes an average out of that. Give samples instead of adjectives.
Common misconceptions
- "The longer the brief, the better." Length does not count, the number of decisions does. A page of adjectives is worth less than three sentences of specifics.
- "Write it once and then reuse it." The audience and the goal are different for every piece. The tone and the banned list, though, really are reusable.
- "More style samples are better." Above five it does not improve, and picking by topic makes it measurably worse.
- "The brief is the editor's job." The brief belongs to whoever carries responsibility for the text. If you do not write it, it will not be your text either.
What to take away
The value of a brief is not in its length but in how many decisions you take in advance — and the most important field is the one most people skip: the specific that only you have. For the tone, five samples are enough, picked at random. Where the topics come from is covered in the Keyword Research with AI post, and how articles add up into a subject area is set out in the Topic Clusters and Pillar Pages post.