Forty people, one room, six hours, a projector that worked on the second try. Here is the honest record.
What stuck
The two-hour build. We spent the whole afternoon on one task: take a real report this team produces every month and rebuild it, badly and quickly, with AI doing the first draft. Everyone left with a thing that half-worked. That half-working thing is worth more than any slide I showed, because they now know where it breaks.
The failure demo. I deliberately ran a prompt that produced a confident, plausible, wrong answer — a citation that didn't exist. The room went quiet in a way that no amount of "always verify" ever achieves. Three people told me afterwards it was the most useful five minutes.
The pairs. Two people per laptop, one driving, one reading the output critically. Solo work produced faster output and worse judgement.
What didn't
The taxonomy slide. The one that maps model types to use cases. It looked impressive and it taught nobody anything, because the distinction it draws does not change a single decision any of them will make on Monday.
The open Q&A at the end. By four o'clock the room wanted to leave. The questions were good; the timing was wrong. I would move it to right after lunch, when energy is high and the build hasn't tired everyone out.
"Any questions so far?" — asked twice, answered with silence both times. People don't ask questions in a group. They ask them at the coffee station, which is where the real curriculum gets negotiated.
The number that mattered
Two weeks later, four of the forty had changed something in how they work. Four out of forty sounds like a failure. It isn't — the baseline for a day of training is usually zero. But if I had designed the day around getting that number to eight, I would have cut the taxonomy slide, moved the Q&A, and spent the last hour on the second build instead of the summary.
Next time.
