AI, usefully · 4 min read · Sep 29, 2026

Give AI a brief you would give a colleague

“Make this better” is a tempting prompt. It is also a surprisingly difficult assignment.

Imagine pasting a page about a proposed product change into an AI assistant. You ask it to make the page better, and it returns something polished: a sharper headline, confident benefits and a neat conclusion recommending a launch.

But the page was meant to help a team decide whether to launch. The awkward paragraph about unresolved customer complaints was there for a reason. The assistant improved the writing while changing the job the document was supposed to do.

If I handed that page to a colleague, I would tell them who will read it, what decision is pending and which uncertainties need to stay visible. AI needs much of the same brief.

Start with the work, not a character

Prompts often begin, “Act as a world-class product strategist.” That may influence the tone. It does little to explain the actual assignment.

A better opening would be:

This is a decision note for a product and engineering review. Help me make it easier to scan. The team has not decided whether to launch. Keep the risks and unanswered questions. Do not add claims that aren’t in my notes.

Now the assistant knows what the document is for. It has a boundary around what it may change.

I think of a useful brief as having four parts:

You do not need to fill out a form every time you ask a question. “Explain this term in plain English” is already a good brief for a small task. The four parts become more useful when the work has consequences or several possible interpretations.

Both Google’s prompt design guidance and Anthropic’s prompting guidance emphasize clear instructions and relevant context. You can test the principle without adopting any special prompt syntax.

Show the assistant what “good” means

Suppose your rough notes concern a new search filter. Early users appear to find results faster, but you have only looked at a small group. Some people also report that useful results disappear when the filter is on.

“Write a launch recommendation” invites the assistant to settle the question.

Try giving it a narrower job:

Turn these notes into a one-page decision brief for a product review.

Use the headings: What we observed, What we do not know, Options, and Next check.

Separate user comments from measured behaviour. Keep the concern about missing results prominent. If my notes do not support a launch recommendation, say so. Do not invent sample sizes, quotes or metrics.

Here are the notes: [paste permitted notes]

This prompt gives the assistant room to organize and clarify. It does not ask it to make the team’s judgment for them.

The requested headings are useful because they make omissions visible. If “What we do not know” comes back empty, you can compare that section against the rough notes. A polished paragraph is harder to inspect when observation, interpretation and recommendation are blended together.

Add an example only when words are not enough

Sometimes you know the result you want but struggle to describe it. A short example can help.

You might show two sentences from a previous decision note: one that reports an observation precisely, and one that labels a possible explanation as a hypothesis. Then ask the assistant to follow that distinction in the new draft.

Use an example of the work, not a long performance of how the assistant should sound. “Clear and direct, with uncertainties visible” usually does more than asking for the voice of a visionary executive.

If the first result misses the mark, tell the assistant where it went wrong: “You turned a customer comment into a measured trend,” or “You softened the main risk.” That feedback gives it something concrete to revise. It does not prove the next version is accurate; you still need to compare it with the source.

Review the brief and the result

Before using AI’s draft, I would check two things.

First, did I provide enough context for the task? If the assistant guessed who the audience was or what “launch” meant, I may need to improve the brief.

Second, did the output stay within the evidence? Check numbers, quotes, dates and any sentence that sounds like a conclusion. If the input contains private or sensitive material, make sure you are allowed to use it with the tool before pasting it.

A good prompt helps the assistant aim at the right job. It does not turn missing evidence into evidence, or remove the need to review a consequential recommendation.

The best improvement may be a shorter answer, a clearer question back to you, or a sentence that says the decision is still open.