# how_it_was_made
The game is inspired by a specific movie scene where a talk-show host fires rapid questions at guests and demands the first word that comes to mind, no time to think. Before any code was written, the project went through a scoping phase: defining the core 60-second loop, deciding what stays client-side versus what needs a backend, writing out a full prompt bank, and working out rules so repeat players don't get the same words twice. The actual build ran largely through Claude Code, handling game logic, UI, backend, and remote multiplayer, with Convex syncing live scores between two phones and OpenAI generating a short AI "verdict" on each round afterward. ElevenLabs voice lines were pre-generated ahead of time so playback during a round is instant rather than a live API call.
// best_dead_end
The AI-written verdicts came out dull or nonsensical for a while, and the cause turned out to be Whisper occasionally transcribing gibberish or fake captions instead of the actual spoken word, which then got passed straight into the verdict prompt without any check.
# prompting_approach
Started with a written scope document answering specific questions before touching any code - who the player is, what the core loop is, what's in versus out for a first version. Once building started, the approach was conversational and iterative with Claude Code rather than one large upfront spec, catching and fixing real bugs (like a live-score race condition between two phones) as actual device testing surfaced them.
