No-code 3D world experiment cycles through four AI coding models
A developer built a 3D world using only , without writing any code by hand. The project started as a 2D zombie game but the focus shifted toward itself. The toolchain rotated across models as usage limits were hit: starting with opencode plus , then switching to claude with fable 5, then opus 4.8, then codex with gpt-5.6-sol, then gemini on auto , and finally back to deepseek once the monthly allowance ran out everywhere else.
Early on, the developer directed the AI to implement a 'decide-evolve-react' pattern, which helped reduce mess in the codebase. Only one line of code was manually fixed, after a model refused to change it — everything else was AI-generated. Every model was very cautious about editing existing code: even when told breaking changes were acceptable, models preferred adding a brand-new function that duplicated an existing one rather than modifying it directly.
This left the codebase with three separate implementations — one using vertices, one using faces, and one using centroids. Most of the codex usage budget went toward refactoring to clean up this duplication.
Key points
- Built a 3D world entirely from , no manual coding
- Rotated through deepseek, claude, codex, and gemini models as usage limits were hit on each
- A 'decide-evolve-react' pattern instruction helped keep the generated code organized
- Models preferred adding duplicate functions over editing existing ones, even when told breaking changes were fine
- Most of the codex budget was spent refactoring away the resulting duplicate BFS implementations