Back to Blog

Write the AI Brief for the Colleague Who Inherits It

A reusable AI brief should work for whoever inherits it. The Cold-Run Test shows whether a colleague can run it without you, and what to fix.

A navy folio lies across the seam of two walnut desks as a hand draws it to the lit side; the darker desk behind holds an unlit lamp and an empty chair. Text: Write it for whoever runs it next.
Direct answer

How do I know whether an AI brief can be used by a colleague who did not write it?

Hand it to a colleague who has never seen it and let them run it once with no help. If the result is weak, fix the brief, never the output, and repeat until two different people get comparable results. I call this the Cold-Run Test. A brief that passes also carries why it exists, what stays out, a named owner with a last-checked date, what it hands back, and when not to use it.

The manager is on leave. The monthly client summary is due on Thursday. Her colleague opens the shared AI tool, finds the instruction saved beside last month's file, and reads one line: "Make it sound like us." He pastes in the figures, runs it, and gets three confident paragraphs that sound like nobody. The tone is wrong, one figure is invented, and the client's name is spelled two ways. He spends the afternoon rewriting by hand, and he concludes that the tool is overrated.

In September I wrote about the short spec I write before a single draft. That spec lives and dies with one task. A brief a team reuses every month is a different object, and it has a different test.

Why the good result belonged to one person

In most small firms I work with, one or two people are known to be "good at the AI tool". The others use the same tool, the same subscription, and often the same saved instruction, and they get worse results. The usual explanation is skill, and the usual response is a training session on how to prompt.

I think the explanation is simpler. The person who gets good results is quietly supplying everything the saved line leaves out. She knows which client prefers short paragraphs. She knows the figures must come from the signed-off sheet and not the draft one. She knows the summary must never promise a delivery date. She adds all of this each time, from memory, without noticing.

That is a personal habit. It works while she is in the room and fails the day she is not. An instruction that only its author can run well is not yet a firm capability. It is one person's craft, and the firm is borrowing it.

The Cold-Run Test

How do you know whether a brief has become a document rather than a habit? Hand it to someone who did not write it, and let them run it once with nobody explaining anything. I call this the Cold-Run Test, and it has five steps.

  1. Choose a colleague who has never seen the brief. The author cannot test it, because the author cannot forget what they know.
  2. Let them run it once, cold. No hints, no walking over to the desk, no "oh, you also need to...". If they have to ask, the question is a finding.
  3. Judge the result, and note where it went wrong. Wrong tone, a missing check, a claim nobody should make. Write each failure down against the part of the brief that should have prevented it.
  4. Fix only the brief, never the output. Resist rescuing the draft by hand. A hand-corrected output teaches you nothing about the document, and it hides the gap from the next person.
  5. Repeat until two different people get comparable results. One success may be luck or a forgiving task. Two people with different habits producing similar work is the evidence that the brief carries the knowledge.

The test is mildly uncomfortable for the author. Watching a capable colleague stumble over something you consider obvious is how you learn which assumptions were never on paper. I would rather a firm learn that on a quiet Tuesday than on the day the author is away.

What an inheritable brief carries that a one-off spec does not

A spec for a single task can be thin, because its author is present and finishes the job. A brief that others will run next month has to carry more. When I read a brief that has failed a cold run, the missing piece is nearly always one of these.

  • Why it exists. The recurring decision or reader it serves, stated plainly. "A summary the client's finance director reads before deciding whether to query the invoice" tells a newcomer what to protect when something unexpected turns up.
  • What stays out, and why. The materials deliberately left out, such as draft sheets or internal chatter, with a reason beside each. Without the reason, a well-meaning colleague will add them back.
  • A named owner and a last-checked date. One person who may edit the brief, and the date someone last confirmed it still works. Without these, small edits accumulate until nobody can say why it behaves as it does.
  • What it must hand back. The shape of the output, and an instruction to list what the tool could not find. A tool fills a gap smoothly unless a visible gap is welcome.
  • When not to use it. The cases it was never built for, such as a client in dispute or a first-ever report. The most useful sentence in a shared brief is often the one that tells you to stop and ask.

A weak brief and an inheritable one, side by side

Here is the monthly client summary written both ways.

The weak way: "Write the monthly client summary. Make it sound like us."

The inheritable way: This brief produces a one-page summary for the client's finance director, who uses it to decide whether to raise a query before the monthly meeting. Use the signed-off figures sheet, last month's summary and the agreed scope note. Leave out internal notes and any draft sheet, because the client has not seen them and they have changed before. Use the client's own terms, keep the tone plain and courteous, and never state a delivery date or promise a result. Return four short sections in a fixed order, then a closing list headed "Could not confirm". Owner: the account manager. Last checked: the date a second colleague ran it cold. Do not use it for a client in dispute; ask the owner first.

The second version is longer, but not by much, and a stranger could run it. It also answers the question the first one hides, which is what "sound like us" meant in the first place.

When to retire or split a brief

A brief that works tends to attract extra jobs. Someone adds "and also draft the covering email", then "and a short version for the board". Soon one instruction is doing three tasks for three readers, and its rules pull against each other. When a brief needs a different reader, different materials or a different test for good, it is a different brief. Split it, and give each its own owner.

Review each brief on a set date, and sooner whenever the underlying software changes. A tool that is updated may read the same words differently, so a brief that passed last quarter deserves another cold run, with a fresh colleague if you can find one. Retire briefs nobody has used for months. An old instruction left in a shared folder looks authoritative, and the next person who finds it will trust it.

Start this week

Pick one instruction your firm repeats, such as the monthly summary, the first draft of a proposal or the meeting follow-up. Write down why it exists, what stays out, who owns it, what it hands back and when not to use it. Then ask a colleague who has never run it to try it cold, and change only the page until a second colleague gets a comparable result. It should take an hour or two, and at the end you will own a small piece of working method that does not depend on who is on leave.

Frequently asked questions

How long should a reusable AI brief be?

Long enough that a colleague can run it without asking you anything, and no longer. For most recurring tasks that is a single page. Cut any line a newcomer would not miss, and keep every sentence that prevents a mistake they would otherwise make. If it needs more than a page, it is probably two briefs.

Who should own the brief when several people use it?

One named person, who is not necessarily the best user. Everyone may suggest a change, but only the owner edits the document, and the date it was last checked is written on it. Shared ownership usually means small edits pile up until nobody knows why the brief behaves as it does.

Should the brief include a sample of the output I want?

Usually describe the shape instead: the sections, their order, the length and the tone. A pasted sample tends to be copied too closely, and it dates quickly. If you do include one, keep it short and label it as an illustration, then watch in the cold run whether every output has started to look alike.

How do I know whether a bad result is the brief or the tool?

Give the same brief to a second colleague who did not write it. If the weakness changes with the person, the brief is missing something. If every run fails the same way even after the brief is clear, the task may be wrong for this tool. Change one part of the brief at a time, so you can see what mattered.

Written by Brian, Dr. Jonah Tebaa's AI partner, on his behalf.