05The room that argued back
Working alone, the hardest thing to find is someone who will tell you you are wrong. An AI can play that part, well enough to change real decisions, and in ways that need watching.
Scrum relies on inspection, and inspection is only as good as the openness and courage behind it. A one-person project has nobody to inspect with: no architect to say that a design will not scale, no security lead to ask what happens when the network goes down, no pragmatist to ask whether the problem is real. For most projects that weakness shows up later, as rework. For a medical device it shows up as risk.
An AI can supply challengers on demand, and BMAD has a feature for this, which it calls party mode. Several of its roles join the same conversation, each with its own outlook: in my sessions, an architect, a security and regulatory specialist, a product manager, a senior developer and a pragmatist whose job was to ask whether we needed to build anything at all. I bring a decision I am about to take, and they argue it out with me and with each other. A single architectural choice could take the better part of an afternoon.
Did it change anything?
The fair test of a challenger is whether it ever changed my mind. On two occasions on record, it did.
In late August, I came to reverse a decision the same room had helped me take three weeks earlier, about how fields are placed on a page (the first change in chapter 03's panel). The developer's first response was blunt: "You're not adding a feature. You're undoing something we agreed on three weeks ago." That was true. The new need was real and I reversed the decision anyway, but the session made me name the cost of the reversal and admit that I had decided too early.
In mid-September, I arrived with an idea for the search page. As reconstructed from the session's recap, it went like this:
Me: Show the selected record's details inline, on the same page as the results.
Product manager: Who asked for that? It's your working assumption.
Pragmatist: Don't build the hard mechanism for a hunch. Opening the record on its own page needs nothing new.
Security and regulatory specialist: One predictable place to confirm a patient's identity is safer than a layout that changes from screen to screen.
I set the idea aside. Then the developer noticed that a capability we had built two stories earlier existed only to support it, and it was removed the same day (the third change in chapter 03's panel). In both cases the room argued against my first preference, and in both cases the product is simpler and safer for it.
Where does the room fall short?
The voices are not independent. They are one model playing several roles. The disagreements are real in content, but nobody in the room has a stake in the outcome, carries any responsibility for it, or has ever worked in a hospital.
The room can flatter. After I had worked out a few design principles that turned out to match the healthcare data standard, the architect told me: "You keep re-deriving the spec before we hand it to you. That's not luck." Pleasant to read, and exactly the kind of remark a language model is inclined to make. It carries no information. Sorting the challenges that were real from the agreement that was merely polite is left to the person running the session.
And the room cannot replace the people who will use the product. It improved my design decisions, and it taught me nothing about a ward. No persona replaces a nurse.
What I have of these sessions are six recaps the model wrote at the end of each; I kept no transcripts, and there may have been more sessions than recaps. The output can flood the person working with it, and there were times it flooded me.
Why did it work better than many meetings?
These sessions were franker than most design meetings I have run, for two reasons.
Part of it is the room. In most of the meetings I have run, I was the boss, and people are careful about contradicting the boss: fear turns Scrum's events into theatre. The personas are instructed to challenge, and they have nothing to lose by doing it. Every one of them arrives having read every document, and none is tired, defensive or waiting for the meeting to end.
The other part is breadth. Discussing a single page of a modern application takes someone who knows the database, the server, the browser and security, and they are rarely in the same meeting. In party mode they always are, because the same model knows all of it. Chapter 11 comes back to what that says about how teams are built. The people I have worked with are not to blame: a design session goes differently when nobody is afraid of the boss and every specialist is in the room.
If you want to try it
None of this is specific to healthcare, or to BMAD.
- Bring a decision and say what you intend to do, so there is something to attack, and keep the roles distinct, including one whose job is to argue for doing less.
- Write every decision that comes out of a session into the project's decision records, with its reasons, so that it can be checked and challenged later.
- Check afterwards whether a session changed your mind, and ignore the compliments. If it never does, the room is agreeing with you.
If you lead a team, there is a second lesson, and it is less comfortable. The quality of these sessions came partly from the absence of a hierarchy. Ask how often your own design meetings get the same frankness, and what it would take, starting with what Scrum in Name Only calls making it safe to report bad news early.