Brief the model like a new starter
Bad AI output is nearly always a bad brief. The team blames the model. The model got one line.
What changed
Facilitators have always been briefers. You brief the sponsor, you brief the room, you brief the test users. What’s new is that there’s a participant in the room who takes briefs literally, has no context you don’t give it, and never asks a clarifying question. It just produces.
Which means the quality of the output is almost entirely a function of what the group managed to write down before pressing enter. And most groups write down one line.
The skill
Turn the group’s shared understanding into a shared brief. The problem, the context, the constraints, what you want back. We use T.O.K.E.N., because people remember it and because each letter forces a question the team usually hasn’t answered yet.
| Means | Example from the sprint | |
|---|---|---|
| T | Task. One sentence. What do you want back, in what form? | “Ten ways to reduce post-bill calls, each as a headline plus two lines.” |
| O | Objective. Why it matters and what good looks like. | “Calls cost £8 each and 40% are avoidable. Good = the customer understands before they ring.” |
| K | Knowledge. Paste the research. | Customer quotes, the journey map, past attempts, what the regulator allows. |
| E | Examples. One you liked and why, one you didn’t and why. | The model learns taste from contrast. |
| N | Nuance. Tone, length, what to avoid. | “Nothing that needs an app. No gamification. Write for a 60-year-old.” |
Fig. 2 — T.O.K.E.N. The team writes it together, on the wall, before anyone types.
A worked example
Same team, same afternoon, same model. First attempt, from a product manager in a hurry:
“Give us ideas to reduce calls after a high bill.”
Twenty ideas came back in eight seconds. A chatbot. An app notification. A bill explainer video. “Proactive outreach.” A loyalty scheme. The team read them in silence, somebody said “we’ve thought of all of these”, and the room’s opinion of AI dropped a notch.
Second attempt, forty minutes later, after we’d built the brief on the wall. The Knowledge section alone was nine customer quotes and a line about the two bill formats that cause most confusion. The Nuance section said no app, nothing that needs a login, must work for someone who pays by cheque.
The ideas that came back were different in kind. Reorder the bill so the comparison to last year is at the top. A single sentence on the envelope. Ring the customer before the bill lands if usage is 30% up. Let the call handler send a one-page explainer by post, during the call. Two of those went to prototype. One is now live.
Nothing about the model changed. The brief did.
What goes wrong
The common failure is the facilitator writing the brief. You’re fastest at it, so you do it, and the team watches. Then the output is yours, and they treat it like yours. Make them write it. Argue over the Objective line. Spend ten minutes on it. That argument is the workshop; the prompt is just where it ends up.
The other one is skipping Knowledge because pasting feels tedious. It’s the section that does the most work. If you have transcripts, paste transcripts. If you have a journey map, describe it. Thin Knowledge, generic ideas. Every time.
Run it Monday
- Put five sticky notes on the wall: T, O, K, E, N. Fill them as a group before any prompt is written.
- Spend longest on O and K. If O takes under five minutes, it’s not agreed yet.
- Run the one-line prompt first anyway, as a control. Then run the brief. Put both outputs side by side. The room will never go back.
- Keep the brief visible all day. Every later prompt inherits it.
If you want better output from the machine, improve the input from the humans.
The filled T.O.K.E.N. above, a blank version and the prompts we run after it are all in the facilitation deck.
