Prompt Brief

Brief an AI tool the way you would brief a skilled contractor on their first day: fast and capable, but with none of your context. Six parts, one test run, one fix, then save it. Works with any capable assistant.

Practice, not research: the six-part structure is working craft, not a tested technique. The risks it guards against, confident false output and personal data leaving your control, are named in the NIST Generative AI Profile, a standard (NIST, 2024).

Use it when:

• You repeat a task with an AI tool and keep rewriting what comes back.
• A task matters enough that you want the same quality every week, not a lucky run.
• You want to hand a working approach to a colleague instead of a vague tip.

Skip it when:

• It is a one-off question with nothing at stake: just ask.
• You are not sure the task belongs with AI at all: sort it first in 10.7 AI Task Map.

This page is practice, not research. No study tells you that a six-part brief beats another structure. The reasoning is simpler: a language model knows nothing about your organization, your reader or your standards unless you put them in the request. Without that, it fills the gaps with the most typical answer, which is why one-line prompts come back generic.

So the brief carries what a capable newcomer would need: context, the task, the limits, the shape of the answer, an example of good, and a rule for what to do when something is missing. That last part matters because generative tools can produce confident statements that are false; the NIST profile for generative AI lists this, under the name confabulation, among its risks, alongside data privacy (NIST, 2024).

The limit: a better brief raises the odds of good output, it does not guarantee it. Anything with numbers or claims still needs checking by stakes. And the brief only gets material you are allowed to share.

Choose a task you do at least monthly. Name the person who reads the result and what makes a result good for them. Then check what material you may give the tool.

Strong answer: "Monthly safety note for our service technicians. Read on a phone between jobs. Good means three changes they must act on, in plain words, under a page." Not: "Write better emails."

The task, and how often you do it:

Who reads the output, and in what situation:

What a good result looks like for that reader:

How long it takes today (rough):

• ☐ The material I will give the tool is public, invented, or cleared of anything personal or confidential. If a task cannot be done without such data, it does not go into a prompt: see 10.3 Rules That Apply to My Role.

Illustrative example: the role, the company and all numbers are made up.

A service manager at an elevator maintenance company in Denmark writes a monthly safety note for technicians. Her first prompt was one line, and the result read like a brochure.

Section 1: reader, technicians on a phone between jobs; good, three actions in plain words, under a page; today about 40 minutes. Material: published notices from the authorities and an incident summary she has already cleared of names, sites and customers.

Section 2: context (service team, field reading); task (summarize this month's changes into actions); constraints (under 250 words, Danish, no claim that is not in the notices); format (three headings: Change, What to do, From when); example (last month's note); when unsure (write MISSING).

Section 3: the first run was generic and listed an assumption that the note was for customers. She sharpened Context only. The second run needed light edits.

Section 4: saved in the team folder as version 2. By the third month the note takes about 10 minutes, and she still checks every date against the notices.