AI
An Agent Is Only as Good as Its Brief: How to Write One It Can't Misread
A vague brief produces confident nonsense. Localization learned long ago that translators need a brief, a glossary and a style guide. Here's what a brief for AI agents needs, with a template you can copy.
On this page
No localization project starts with "translate this, thanks." It starts with a brief: who the text is for, where it will appear, what tone to use, when it's due. Next to the brief sit a glossary of approved terms and a style guide. Skip them and you still get a translation. It's fluent, it's consistent with itself, and it's wrong for the client.
AI agents fail in exactly the same way, only faster. That's why, in every run I coordinate, the brief comes first, and nothing runs until the requester has approved it.
Why a prompt isn't a brief
When a person is missing information, they usually ask. An agent usually guesses. It fills every gap with the most plausible assumption, then builds on that assumption with complete confidence. A vague brief doesn't produce a vague answer. It produces a precise answer to a question you didn't ask.
A prompt says what to do. A brief also says what success looks like, what's off-limits, what the words mean and how the result will be checked. Its job is to leave nothing important to guesswork.
The six parts I won't run without
1. The outcome, in the requester's words
Not "build a reporting tool", but the result the person actually needs: "one accurate report by 10:00 every Monday, run by the team without a developer." Every option will be scored against the outcome later, so it has to be concrete enough to fail.
2. Constraints: what must not happen
Constraints do more work than goals. "No client data uploaded to new services" and "no silent errors" rule out whole families of solutions before anyone spends time on them. Write down the things that would make you reject an answer, however good it looks otherwise.
3. This agent's angle
When several agents explore the same problem, each one gets a different job: one researches existing tools, one designs the simplest thing that could work, one designs for the messiest input you have. Same outcome, same constraints, different angle. Without that, you get five versions of the same first idea.
4. Inputs and data rules
Which files, documents and systems the agent may use, and which it may not. Where possible, give it anonymized or synthetic samples rather than real files. In localization this is second nature: confidential material goes only where the NDA allows. Agent work deserves the same discipline, agreed in writing before anything runs.
5. What to hand back, and the evidence
Say exactly what the result should look like: a comparison table, a one-page design, working code with tests. Then say what counts as proof. My standing rule is that every claim needs evidence: a source, a test run or a sample output. Anything the agent couldn't verify must be labeled as unverified, not dressed up as fact.
6. A glossary
Every organization has words that mean something specific. "The report" might mean the Monday client report, not any report. An "error" might mean a wrong figure, not a formatting slip. "Done" might mean "approved by the client", not "sent". A translator would get a termbase for this. An agent needs one even more, because it will never ask.
A template you can copy
BRIEF: <short name>
Outcome
What the requester needs, in their words. Concrete enough to fail.
Constraints (reject any option that breaks one)
- ...
- ...
Your angle
Your job in this run, and what is NOT your job.
Inputs and data rules
May use: ...
Must not use, copy or send anywhere: ...
Deliverable
Format, length, and what to include.
Evidence
Every claim cites a source, a test run or a sample output.
Label anything you could not verify as UNVERIFIED.
Glossary
<term>: what it means here.
Open questions
List everything you had to assume. Do not guess silently.
The last section matters more than it looks. Asking the agent to list its assumptions turns silent guesses into questions you can answer.
Briefs for the challengers
The agents that review the results get a different brief. Their inputs are the original brief and one result. Their instruction is to find where the result fails the brief: wrong facts, missing cases, broken constraints, assumptions nobody stated. They're told not to fix or rewrite anything, only to report the problems, ranked by how much they matter, each with its evidence.
That separation is deliberate. A reviewer asked to improve a draft polishes it. A reviewer asked to break it finds what's wrong. (More on that in Nobody Reviews Their Own Work.)
Signs your brief is too vague
- You would accept two very different answers.
- You can't say in advance what would make you reject the result.
- Two runs of the same brief come back with very different results.
- The answer surprises you with an assumption you never stated.
- The words you'll judge the result by ("simple", "robust", "fast") appear nowhere in the brief.
Any one of these means the agents will decide something you should have decided.
Approve the brief, then run
The brief is the cheapest place in the whole process to fix a misunderstanding. Correcting a sentence in a one-page brief takes a minute. Correcting the same misunderstanding after the agents have built on it takes a day, and correcting it after delivery costs a lot more than that.
Translation teams learned this long ago: answer the query before translating, not after the client spots the mistake. That's why the requester approves the brief before any agent runs, and why the same brief is the yardstick the final result is checked against.
See how the brief fits into the rest of a run: AI agent orchestration.
About the author
Written by Ioannis Ouzounis, an independent AI and automation consultant. He coordinates teams of AI agents to explore, test and build solutions, and personally signs off on every result.
More about Ioannis