The Briefing Playbook

The most common complaint about AI agents is that the output is generic. In almost every case we have looked at, the output is exactly as specific as the brief was, and the brief was a topic.

A topic is "write a LinkedIn post about our new pricing". A brief is a topic plus four other things. This playbook is those four things, some templates, and the mistakes that recur.

The five parts

1. The decision already made

Start with what is settled, because everything you leave unsettled will get settled by the agent, silently, inside the draft. "We are leading with switching cost, not price" is a decision. "Talk about our advantages" is not.

This is the part people skip because it feels obvious to them. It is not obvious to anything that has not sat in your meetings.

2. Who it is for, named tightly enough to exclude

Not "B2B marketers". "Operations directors at hauliers running five to fifty trucks, who own the dispatch board personally." The test is whether the description excludes someone. If it does not, the output will be written for everyone, which is a well-understood way of being read by nobody.

3. What the reader should do or believe differently

One sentence. If you cannot write it, the piece has no purpose and you will not be able to tell whether the draft succeeded. "Ask their dispatcher what percentage of last month's miles ran empty" is a good one. "Understand our value proposition" is not, because there is no way to fail it.

4. The constraints

The half of a brief that is almost always missing, and the half that does the most work.

• Claims you may not make. Anything unproven, unapproved or legally awkward.
• Numbers. Either supply them with sources, or say plainly that no number may be invented and an unsourced claim should be flagged rather than filled.
• Length and format, as a real constraint. "Around 250 words" produces a different piece to "as long as needed".
• Voice, expressed as things to avoid rather than adjectives. "Avoid opening with a rhetorical question" beats "punchy and confident" every time, because the first is checkable.

5. What good looks like

One example of the standard, or one example of the failure. A single previous piece you were happy with does more than a paragraph of description. If you have nothing to point at, describe the failure instead: "the last version read like a press release, which is the thing to avoid."

Template: content

• Decision made: we are arguing [position], not [rejected alternative].
• Reader: [role] at [company type], who is judged on [metric].
• After reading, they should: [one concrete action or changed belief].
• Must not: claim [x]; cite any number without a source; use [format or opening to avoid].
• Good looks like: [link or paste of a previous piece], particularly [the specific thing it did well].

Template: outbound

• Trigger: the observable event that makes now the right time to contact this person.
• Reader: [role], who currently believes [assumption] and has probably [prior experience with this category].
• The ask: exactly one, and it is not "a quick call" unless a call is genuinely the next step.
• Must not: open with a compliment; reference the trigger without saying why it matters; mention a feature before the second email.
• Stop condition: if [signal] after [period], the trigger is wrong and the copy should not be rewritten.

The stop condition is the part nobody writes and everybody needs. Without it, a failing trigger gets diagnosed as failing copy, and the team rewrites emails for a month.

Template: strategy

• The question: stated so that a wrong answer is possible.
• What is already decided: [constraints that are not up for debate].
• What is genuinely open: [the actual decision space].
• Evidence available: [what exists, and where]. Say explicitly what does not exist.
• Must not: recommend anything requiring [resource you do not have]; present an option without naming its cost.
• Deliverable shape: a recommendation with a named risk and what would change your mind.

Six anti-patterns

Briefing a topic. The default failure. Fix: add the decision already made.

Asking for "best practice". Best practice is by definition the thing everyone is already doing, so it produces output that fails the competitor-swap test. Fix: ask for a recommendation for your specific constraints, and say what they are.

Demanding authority without supplying sources. "Be authoritative, use data" with no data attached is a direct instruction to fabricate. Fix: supply the evidence, or say that unsourced claims must be flagged rather than filled.

Describing voice with adjectives. "Confident, warm, human" cannot be checked and so cannot be met. Fix: two or three concrete prohibitions.

Leaving length open. Unbounded length reliably produces padding, because length is the easiest proxy for effort. Fix: state a range, and mean it.

Reviewing by rewriting. You fix one artefact and teach nothing. Fix: name what failed and send it back once. The second draft is usually better than your edit, and the brief improves permanently.

Pre-send checklist

• ☐The brief names a decision, not just a topic
• ☐The audience description excludes somebody
• ☐There is one sentence describing what should change for the reader
• ☐Unmakeable claims are listed
• ☐Numbers are either supplied with sources or explicitly forbidden
• ☐Length is bounded
• ☐Voice is expressed as prohibitions, not adjectives
• ☐There is one example of the standard, or of the failure to avoid

If six of these are ticked, send the brief. If three are, you are about to spend longer reviewing than you saved.