One Friday evening in March we had the first working version. Lobby worked. Contributions went through. Reveal was messy but readable. Time for a real test.
Setup
Four friends, all in the same small office. Everyone on their own laptop. Meant to wrap up in an hour. Became four hours.
What broke
- The WebSocket connection dropped when someone closed their laptop. No reconnect logic. We had to kick them from the game and have them rejoin.
- An AI bridge came back in English. Not easy to read, and not remotely funny.
- The ending took 30 seconds to generate because we'd accidentally configured Claude Sonnet instead of a fast model. That's where we learned about latency.
- The timer ran out before someone could submit because they were still typing. We've since added a "currently typing" status.
What worked
The sentences. The chain sentences were hilarious from the very first moment. One player wrote "The dog was too heavy for the swing," the next "The swing blinked once and ducked." Nobody had expected the swing to blink. Everyone laughed.
At the moment the reveal started, and we read together the three stories we'd written in parallel, something clicked into place. It was no longer a prototype that technically worked. It was a game that was fun to play.
Bug list that night
I wrote a list of 23 bugs and ideas in my phone that evening. Of those, 21 have now been fixed, weeks later. The last two are "nice to have" but not critical.
A prototype is always ugly. But when the core works, you can build on it. That evening we knew: the core works.