Recent advances in large language models have shown remarkable abilities in web front‑end execution, making browser‑based game generation a prominent frontier.
Previous works often rely on complex multi‑turn workflows or evaluate only static games, whereas this study focuses on end‑to‑end real‑world game synthesis driven by coding agents.
Generating complex games directly from brief user queries forces agents to make underspecified assumptions, leading to incomplete mechanics, disconnected gameplay, and limited visual quality.
To address this, we introduce GameGo, a scalable framework that systematically converts short game seeds into industry‑standard Product Requirements Documents (PRDs). It preserves core gameplay constraints while employing task‑specific dynamic compression to maximize information density without restricting design exploration.
Based on this pipeline, we build GameGoData, containing 55,060 development trajectories across 2D, 2.5D, and 3D games, and release GameGoBench, a benchmark with 124 diverse game queries.
Training GameGoCoder on GameGoData yields a model that outperforms matched baselines and rivals frontier models on gamedev benchmarks.
All code, datasets, and models will be publicly released.
Review: GameGo demonstrates that compressing brief ideas into industry‑grade PRDs can provide high‑density, instruction‑following signals, offering a practical and scalable route for LLM‑powered game development.