The mock meeting and the one-page brief

False beliefs first. Any number above zero means a leading question slipped through, and each one is a build decision you will get wrong in week six. On this bank the costliest is "the assistant replies to suppliers on its own": it is the vendor chatbot again, and it is wrong because only the area manager may say yes to a substitute. Find the leading question in the transcript, write the open version of it on your sheet ("In a thread, which replies need a person to decide, and which are just looking something up?"), and carry that version into the real meeting.

Doors opened. Six is the only answer that lets you write a brief without guessing. A shut door is a block of the brief you cannot fill. For an AI developer the door that matters most is door 3: if you never walked one thread end to end, block 1 has no workflow in it and block 3 cannot say where each half of the thread lives.

Net value against the ladder. The Six Doors in printed order score 54 blind; tuned for an AI developer, 55. That one point is honest and small. The printed order already asks the lookups-versus-decisions question, a decision-door question that does the current-way work, by minute 12, and it hears the one-in-twenty error rate at minute 16. A run in that region means the sheet works. A run near the shuffle's mean of 30.1 means the questions are good and the order is not: usually the thread was never walked, or jargon crept in early and cost trust the rest of the meeting needed.

The 20-minute counter. The tuned order puts 23 points in your notes by minute 20, printed order 22, a shuffle about 14.3 on average. If she is called away halfway, that is the meeting you actually had. The tuned order's 23 already holds the thread, the error rate, the lookup-or-decision line and the worst case, which is most of a block 1 and a block 2.

The tuned run asks eleven questions in 45 minutes, opens all six doors, and still leaves six facts on her side unsaid: that the board wants the word AI on a slide; who holds the POS back-office login, the most sensitive fact on the bank, which needs the most trust before she says it; that customer names sometimes appear in the threads; that anything the managers use has to live inside WhatsApp or on the phone, in the van; that the board meets in six weeks and the supplier contracts renew in three months; and that Farid will trial anything and Priya in finance can give two hours a week. None of that is a failure of the sheet. It is the shape of 45 minutes. What matters is what you do with it: every one of those is a line the brief marks "to confirm with you" or a question in the covering note, and never a sentence written as if she had said it.

The bench scores what you asked and what came back. It does not hear your tone, see whether you looked at her or at your laptop, or notice the moment she leaned forward because a question was about her managers' actual Tuesday. Those matter, and no widget measures them honestly. The run is a rehearsal for the order and the kinds, not a verdict on you as an interviewer. If you can, run the real meeting with a colleague taking notes, and ask them one thing afterwards: when did she start giving longer answers?

1 · What you told me, in your words. Two or three sentences she said, quoted, and at least one of them is the thread as she walked you through it. Not your summary. "My three area managers spend half their day on them" is her sentence; "high manual message-handling load" is yours, and she cannot correct a tune she never hears. This is the block that closes the gap from session 1: if what you heard is not what she meant, she says so here, and it costs one email instead of six weeks.

2 · Which replies are lookups, and which are decisions. Session 4's one concept, written as a fact about her shop. Inside a thread one kind of reply needs a person, accepting a substitute or a price change, and only the area manager may say yes to it; everything else is looking something up. Put the error rate today beside it: about one confirmation in twenty carries a wrong price or a substitute, and the managers catch most of them. That number is the bar any assistant has to beat, and how they catch it is the part the assistant has to keep.

3 · Where each half of the thread lives. The thread map from session 2, as a thumbnail or six lines of text. A thread has two halves and they live in different places: the supplier's confirmation in a chat on a manager's personal phone, the outlet's original order in another WhatsApp group, in three languages with nicknames. End it with one question: "which of these is wrong?"

4 · What I will build first, and what I never will. The pilot from session 5 in three short lines, in its order: the decision log first, the confirmation format at the next contract renewal, and only then an assistant that drafts the plain-OK replies for one manager and four suppliers, shows the match it made, and never approves a substitute or a price change. Then one sentence naming what it will not do. "The assistant will never approve a substitute" is the line that most needs her agreement, because it is her door 2 answer written as a specification, and because she has already switched off one system that answered on its own.

5 · How we will both know. Her test, in her words: the area managers visit two outlets a day, and she reads the visit log herself. Not an accuracy figure. She told you she does not care what the assistant's accuracy number is, and a brief that leads with one tells her you were not listening.

6 · Who does what next week. Names, a day, a time, one task each. Rob Fitzpatrick's test in The Mom Test (2013, reported from summaries) for whether a conversation went anywhere is whether the other side committed something: time, a name, an introduction. This block is where her commitment is written down so it cannot quietly lapse, and where anything the meeting never reached is marked "to confirm with you" instead of being filled in.

Within a day she still remembers what she meant, so a correction is cheap and specific. After a week she remembers that you came by, and a brief arriving then reads as a proposal to be evaluated rather than a record to be checked. No study is quoted for the 24 hours; it is a working norm, and the reason is the one above.

One page because the people who act on it read on their phones: her managers are in the van, and she reads between outlets. Two pages is a document; a document goes into a folder to be read later; later is after the board. If a block needs more than three lines, the detail goes in an attachment she can ignore.

A word from the technical opener: fine-tune, retrieval, hallucination rate, eval set, vector store, guardrails. An accuracy figure for an assistant that does not exist yet. A customer's name, or a quoted message that carries one: a patient name in a message makes it personal data under Malaysia's PDPA, and the head office lawyer takes four weeks to answer anything, so the brief says "no message with a customer's name in it goes into phase one" and shows none. A sentence she never said, written as if she had; the six facts the run did not reach are questions, not statements.

And a fact she told you only once trust was high. If your run reached the question about who can log in to the POS back office, and she told you what she has not yet told her team, that fact shapes block 4, because the step that needs the product list must not depend on one login, but it goes into your plan and not into a document she may forward. The brief is hers to forward. Write it as if it will be.