On brand, on time

My last role came with a standing delivery: training for the whole company, every week, at a quality bar that never moved. Not only sales training either: sales, operations and product, because a seller who knows the pitch but not how an order actually gets fulfilled will still lose the account three months in.

Voice Media Group ran several divisions selling largely the same products into different markets, so the material had to work everywhere without being rewritten for each one. I was the person producing it.

Weekly output against a fixed standard fails in two ways, and both are visible from outside. The date slips, or the material lands thin. Nobody notices a production system working, but everybody notices a training that shows up late or shallow. The bar was the one thing I could not move, so production had to change instead.

I built a multi-agent system to do the production - one orchestrator, three subagents, taking a subject from research through to a finished deck. Then I spent most of the effort testing it, running roughly twenty cycles before it made anything anyone saw.

Every cycle worked the same way. Run it, compare what came back against the spec, then hand both the instructions and the output to a fresh pass and ask where the two diverged. Whatever that turned up got written into the spec, and the next cycle started from there.

Handing back both halves is what made it work. Output on its own tells you something came out wrong. Put it next to the instruction that produced it and you can see which line to change.

Two things came out of the cycles that we would not have found by using the tools normally.

The deck tool had a ceiling: past about twenty-one slides it stopped including content, with no error and nothing on the finished deck to show that anything was missing. I capped the slide budget below the ceiling and split longer subjects across sessions.

The other one was worse: the assistant we used for research was reading the entire workspace whenever files were not explicitly scoped to it. In a company with client material sitting in shared drives, that is the kind of default nobody finds until it costs them something. Scoping became a required step.

Neither behaviour is documented anywhere, and you find that sort of thing by running the same test twenty times and reading what changed.

The decks were usable on arrival. They matched our voice, our visual system and the conventions the training program already ran on, because I wrote those into a spec the system generates against rather than correcting them by hand afterward. It ran on tools the company was already paying for, so putting it in cost nothing.

The other half of the design is where a person sits. Every statistic has to carry a publisher, a date and a URL, and anything without one gets tagged and cannot pass. On contradictory sources the system stops and escalates rather than picking a side. Nothing ships until a person signs a source ledger.

None of that is a safety net bolted on afterward. It is where the design puts the judgment, on purpose. The machine does the production and a person decides anything that requires deciding, and the points where control changes hands are specified rather than left to whoever happens to be paying attention that week.

That is also why it could be handed over. If the judgment had lived in my head, nobody else could have run it. Judgment is a step in the process with an artifact attached, so my successor could take the same seat.

It produced the last two trainings I delivered, and I demoed it live to my successor, who kept running it after I left.

I designed around the slide ceiling and I was pleased with myself for it. The better move would have been to ask whether that was the right tool to build on at all.

Engineering around a limitation feels like progress, and sometimes it is. But a ceiling you have to route around is information about the tool, and I treated it as a puzzle instead of a signal. Next time I would spend the first cycle choosing the platform rather than the twelfth working around the one I had already started with.