A new research pipeline turns a one-line game idea into a full production document, then uses that document to train AI coding agents on tens of thousands of real builds.
The system, called GameGo, comes from a paper posted to arXiv on October 7. Instead of asking a coding agent to build a browser game straight from a vague prompt, which tends to produce broken mechanics and thin visuals, GameGo first expands that prompt into a detailed Product Requirements Document modeled on real game-studio practices. A compression step keeps the document dense enough to follow without boxing in the design. The team used this pipeline to generate GameGoData, a set of 55,060 development trajectories covering 2D, 2.5D, and 3D games, plus GameGoBench, a 124-query test set, and trained a coding model called GameGoCoder on the results.
The underlying problem is familiar to anyone who has watched an LLM try to one-shot a game from a sparse prompt: it fills gaps with guesses, and the guesses are usually generic. GameGo's bet is that the fix is not a bigger model but better intermediate structure, forcing the agent to plan like a studio would before it writes a line of code. According to the paper, GameGoCoder beat similarly sized baselines and came close to larger frontier models on gamedev benchmarks.
That comparison comes from the same team's own benchmark, so treat "comparable to frontier models" as a claim to retest once the promised code, data, and models actually ship.