← all posts site ↗

my game idea miner now writes design docs

there is a twitter post i could not find but remembered perfectly. ai cannot create new concepts, so someone scraped a whole storefront catalog, embedded it into a high dimensional space, picked a mathematically distant point in the cloud, and had a language model convert that point back into an idea. the best comment said the obvious thing, new does not mean valuable, most of those games would be different but bad. someone else suggested running the same trick in gameplay space instead of concept space, and that is the version i actually wanted.

so the project grew a cli. one command prints a random idea. another walks the map by tags. another finds the nearest neighbors of a known game, and a table command dumps the whole catalog as tsv so it pipes into anything. sixty two thousand games in the map, thirty eight thousand of them with quality metrics. the neighbors output is the part that feels like a superpower, asking for one game and getting its five nearest relatives with distances.

the output shape tightened into a real design doc. one liner, art direction, gameplay verbs, end state, tags. then a feedback round added goal, play, fail state, run length, scope, and the best addition, parallels. heavy combat like a souls game with lane movement like a runner. that row is what makes the idea click in your head. the caveat is honest, the reference games come from the model's training, not the corpus, so if a parallel is wrong, that is the model, not the data.