The tool ran on an iPad in someone's hand, in a room where a decision about a house was being made. Nobody was watching it perform. They were using it, moving things around, changing their mind, deciding whether a space worked, while a real showing happened around it.
That's the line I care about. Not that it was built, not that it demoed well. That it got used, live, by a person who didn't care that it was a tool, in a moment that mattered to them for reasons that had nothing to do with the tool.
A demo and a delivery look almost identical right up until the moment of use, and then they diverge completely. A demo is built for the person evaluating it. Every choice serves the impression, it has to survive being watched, being judged, being asked "so what does it do." A delivery is built for the person using it, who has already stopped evaluating and started doing the actual thing they came to do. They're busy.
The configurator was the second kind. It ran inside a Claude app artifact, no install, no setup, no "let me get this working first," just a link that opened into a working tool on a tablet. It let someone stand in a real space and rearrange it, try the couch on the other wall, see whether the room they were considering could become the room they wanted. The AI part was invisible to them, which is the point. They weren't using an AI tool. They were moving furniture and making up their mind.
Demo conditions flatter everything. In a demo, the WiFi is fine, the inputs are the ones you prepared for, the person is paying attention because paying attention is their job in that moment. Real conditions flatter nothing. The room is a real room. The hand holding the iPad belongs to someone with an actual decision in front of them and no patience for a tool that gets in the way of it.
A tool that only works in the demo has been rehearsed, and that's all. The gap between rehearsed and finished is exactly the gap that real use exposes, and the only way to know which side of it you're on is to hand the thing to someone who has a reason to use it that has nothing to do with proving it works.
Nobody clapped. Somebody moved a couch on an iPad, decided they liked the room better that way, and kept going. That's the whole test. Not the applause, the couch.
[SOURCED] A room-configurator tool built for a real-estate client, running inside a Claude app artifact on an iPad, used live during an actual showing.
• Client kept to a vertical descriptor only, no name, brokerage, location, or date, by rule.
• The specific outcome of the showing is not asserted; the claim is that the tool was used in real conditions, not that it closed a sale.