Prompts & Writing
Writing Briefs That Produce Good Output, From Humans or Models
By Jim Vernon, Editor, AI Intelligence International · Published 8 February 2026 · Reviewed against our editorial standards · About the author
Teams adopting AI often discover their briefing process was never good enough. It survived because experienced humans quietly filled the gaps from context. A model does not fill gaps; it invents plausible ones.
Fixing the brief improves human output too, which makes it the highest-leverage process change in most creative and analytical teams.
Key takeaways
- The six elements: Audience and their current belief; the single change you want in that belief; the evidence available; the constraints; the format and length; and the definition of done.
- Definition of done: Write the acceptance criteria before work starts: what must be present, what must be true, who signs off, and what would cause a rejection.
- Evidence supplied, not implied: Attach the data, the quotes, the previous version, the competitor page.
- Constraints that matter: Legal and regulatory limits, claims you cannot make, competitor mentions, brand vocabulary, accessibility requirements, and anything the last version got wrong.
The six elements
Audience and their current belief; the single change you want in that belief; the evidence available; the constraints; the format and length; and the definition of done.
Most briefs contain topic and deadline only. That is not a brief, it is an assignment.
The 'current belief' element does the most work and is almost always missing. Writing changes something; you cannot change something without naming its starting state.
Definition of done
Write the acceptance criteria before work starts: what must be present, what must be true, who signs off, and what would cause a rejection.
This converts subjective review into a check, which shortens revision cycles dramatically.
It also gives you the rubric you need if the work is later model-assisted.
Evidence supplied, not implied
Attach the data, the quotes, the previous version, the competitor page. If the executor has to hunt for material, the output will be built on whatever is easiest to find.
For model-assisted work this is decisive: unsupplied evidence becomes invented evidence.
If no evidence exists, say so explicitly and commission the research as a separate step.
Constraints that matter
Legal and regulatory limits, claims you cannot make, competitor mentions, brand vocabulary, accessibility requirements, and anything the last version got wrong.
Constraints are cheaper to state than to correct. Every constraint omitted is a revision cycle.
Keep a standing constraints block that applies to all briefs so you only write the exceptions.
Length and format decided upfront
Length should follow from the job, not from habit. Decide the format — page, email, deck, memo — and the approximate length before anyone starts, because both change the argument structure.
Requests to 'make it shorter' after delivery usually mean the brief chose the wrong format.
A reusable template
Audience and current belief. Desired change. Evidence attached. Constraints. Format and length. Definition of done and approver. Deadline.
Seven lines, five minutes to fill in, and it works identically whether the executor is a junior, an agency or a model. Run your next three briefs through it and compare revision counts against the previous three; the difference is usually obvious.
The six fields that do the work
Audience and what they already know. The single outcome you want. The facts and sources that must be used. Hard constraints such as length, format and prohibited claims. Tone, with one example. And who signs it off.
Those six fit on half a page and remove almost all the back and forth. Briefs fail from omission far more often than from length, and the most commonly omitted field is the outcome.
Write the outcome as a sentence the reader should be able to say afterwards. 'The reader knows which plan to choose and why' is a brief; 'raise awareness' is not.
The same brief works for people and models
A brief good enough for a freelancer is good enough to paste into a prompt, and vice versa. That is a useful test: if you would be embarrassed to send it to a contractor, the model will also produce something disappointing.
The difference is that a contractor asks clarifying questions and a model does not. So add one instruction models need and humans do not: list any assumptions you had to make before writing.
That single line surfaces the ambiguity your brief left behind, and it is faster than reading the draft to discover the same thing.
Brief once, reuse forever
Recurring work deserves a stored brief rather than a fresh one each time. Monthly reports, release notes, case studies and customer updates all have stable requirements that change perhaps twice a year.
Keep the stored version next to the last approved output. The pair — brief plus accepted result — is the fastest onboarding material any new person or new tool can be given.
Update the brief whenever feedback repeats. A second identical revision request is a brief defect, not a writer defect.
Frequently asked questions
Is a template too rigid for creative work?
The template constrains the problem, not the solution. Creative teams generally report faster approval because disagreement surfaces at brief stage instead of at delivery.
Who should write the brief?
Whoever owns the outcome, with input from the executor. Briefs written by the executor tend to encode what is easy rather than what is needed.
How detailed should evidence be?
Enough that no fact needs to be invented. If you cannot supply it, mark it as an open question rather than leaving it silent.
Does this apply to code work too?
Yes, almost unchanged: audience becomes the calling system, desired change becomes the behaviour, definition of done becomes tests.
How long should a brief be?
Half a page for most work. If it runs past a page, the piece is probably two pieces or the requirements are not yet decided.
Who should write the brief?
Whoever will approve the result. Briefs written by someone without sign-off authority get overruled at the end, which wastes the whole cycle.