A game developer explains how to make a project legible to AI coding agents

Kyle Kukshtel turned his conference talk "Your Game Is A Harness" into a long post. His point: LLMs struggle with games because so much of a game is not code, so the fix is better context, not a better model. His tips: use one seeded source of randomness so any game state can be reproduced; let the agent take screenshots and compare them with an image-diff tool (he uses odiff), which gives a similarity score it can work towards; use one simple object type and build save and load from day one; make every change a command that can be replayed; keep game data outside the code so tweaks need no recompile; and run the game's simulation separately from the engine. On Hacker News (about 85 comments), opinions split. Some said it matches their own experience, especially screenshots and a separate simulation. Others said these are simply good engineering habits, or that the advice suits small indie games but not large 3D games with tight performance budgets. A few game developers argued coding was never the hard part of making games; art and design still are. Sorted take: the lesson travels well beyond games. An agent does better when it can reset to a known state, check its own work with a number, and change settings without a rebuild. Ask what that looks like for your own app.