How to give AI a useful brief: goals, context and clear constraints

If you asked a new colleague to "write a marketing plan", they would ask questions. For whom? How much money? What do we already know? When is it due? A language model usually does not ask. It fills the gaps with the most plausible choices, and you receive something that looks finished and may be wrong for your situation.

A brief is how you supply those answers up front. It does not need clever wording. It needs the right information.

Why vague requests disappoint#

"Write a marketing plan" leaves almost everything undecided: the business, the goal, the audience, the budget, the channels, the tone and the format. The model picks defaults. The plan it produces may be reasonable in general and useless for you.

Anthropic's prompting guidance puts the principle simply: be clear and direct, and add context that explains what you are trying to achieve. Think of it as briefing a capable colleague who has no background on your business.

Six parts of a useful brief#

1. Goal. What should exist at the end, and what is it for?
2. Audience. Who will read or use it?
3. Context. The facts the model needs, and only facts you are happy to share. Separate what is true from what you want suggested.
4. Constraints. Budget, length, tone, things to avoid, things it must not claim.
5. Output. The structure you want: sections, a table, a short list.
6. Unknowns. Tell it what you do not know, and invite it to ask or to flag gaps.

A before-and-after example#

The business below is invented.

| | Request | | --- | --- | | Before | Write a marketing plan. | | After | The brief below. |

Goal: Draft a 6-week marketing plan to fill a new Saturday pottery class at Clay & Co., a small studio.
Audience: The plan is for the studio owner. The class targets adults in the neighbourhood who have never made pottery.
Context (approved facts): Class runs Saturdays 10:00 to 12:30. Price is 45 per person. Eight places. Studio has 1,200 newsletter subscribers and an Instagram account.
Constraints: Budget 300 in total. No paid influencers. Do not promise results or invent testimonials. Friendly, plain language.
Output: A week-by-week table with action, channel, cost and what to measure. Then a short list of risks.
Unknowns: We do not know our current conversion rate or how many past students rebook. List what you would need to find out, and mark any assumption clearly as an assumption.

This brief does five things the vague request did not. It names the goal and the audience. It supplies approved facts and the budget. It limits claims. It asks for a structure. And it says what is unknown, so the model should flag gaps instead of filling them with invented figures.

Ask for gaps and uncertainty#

The final line matters most. A model that is allowed to say "I do not have enough information" is less likely to produce a confident invention. You can add:

• "If something is missing, ask me before you answer."
• "Mark anything that is an assumption."
• "If the provided text does not contain the answer, say so."

Documentation on reducing hallucinations makes the same point: giving explicit permission to admit uncertainty can substantially reduce false information. It reduces the problem. It does not remove it, which is the subject of the next article.

Improve with specific feedback#

The first reply is a draft. Respond like an editor, not like a judge:

• Say what to keep: "Keep the week-by-week table."
• Say exactly what to change: "Replace week 3 with something free. We cannot spend more than 50."
• Add the missing fact: "Our conversion rate is about 4 percent. Redo the forecast."

Specific feedback gets better results than "make it better".

Know what not to paste#

A good brief contains facts, which means it can contain private ones. Before you include a customer's details, a contract or an employee's record, check that you are allowed to share it with this tool and where it is processed. When in doubt, describe the situation with a made-up example instead of the real data.

Review before you use it#

Even a well-briefed answer is a draft. Read anything with names, numbers, dates, promises or legal and medical statements against a real source. A brief improves your odds. It does not replace checking.

There is no magic prompt#

You will see lists of "perfect prompts". The honest version is smaller: say what you want, why, for whom, with which facts and limits, in what shape, and what you do not know. Different tools respond a little differently, so adjust when a result misses.

Next: Why AI can sound confident and still be wrong, and how to check what it tells you.