The AI writes the game code on its own. The art is getting more and more complicated.

A developer single-handedly created a game in two days—entirely using his own o2 engine, with code from *Claude Fable 5* and art from *Nano Banana 2*. The most interesting part here isn’t the fact that it worked, but rather where the AI handled things easily and where workarounds had to be built on top of it. Spoiler: the code isn’t a problem; the art is hell. All of the game’s code was generated by Claude Fable 5. The author reviewed the code but didn’t write it himself. The key to success isn’t the model itself, but the infrastructure surrounding it. In the engine’s repository, alongside the source code, there’s Markdown documentation describing the architecture and subsystems. The rules for the AI state: before reinventing the wheel, check the documentation. The engine is only a few megabytes in size, which also helps: the context fits without any clipping. The second factor is the tests. The engine includes a test driver: screenshots, clicks, and frame scrolling. The AI writes code, runs the test, and displays the result. The project’s basic rule is to write tests for everything and generate a report. This turns the work into a game design dialogue: “Make the car bounce slightly after a turn”—and the AI goes through the process on its own: coding, testing, and displaying the results. We’ve accumulated over 50 screenshot reports that show how the game is “growing.” It’s worth pausing here to be honest: code generation works when you already have documentation, a test driver, and a small project. This isn’t “AI will replace programmers”; it’s “AI speeds things up for those who already have a process in place.” With art, it’s a different story. The main problem isn’t the quality of a single image, but consistency. You can generate an icon, a button, or a house—individually, they look fine. But in the game, they don’t fit together: each time, the AI starts from scratch and does whatever it sees fit. The author put together a node-based tool for generation pipelines: you define a concept, and screens, objects, and sprites are derived from it—all with a reference to the original concept. This solves part of the problem, but by no means all of it. An isometric city map comes with strict constraints: a grid, angles, and precise tile alignment. The generator “drifts”: tilt, perspective, shadows. We had to build a deterministic pipeline. A script generates line art for buildings with correct isometric projection → feeds it with a prompt into Nano Banana 2 → the script analyzes the result, rejects pixels that are out of place → sends them for re-generation → final validation by a human. For road tiles, it’s even stricter: 2D vector generation → conversion to isometric → programmatic application of textures and markings → running through Nano Banana 2 with the prompt “improve, but stay within the boundaries.” An attempt to generate entire city blocks using AI failed: even a highly deterministic pipeline yielded at most 5% good results. Shadows were distorted, isometric projection broke down, and the block spontaneously grew upward. The idea was scrapped. Sound familiar? This is exactly the same problem faced by everyone trying to use generative models in production: a single generation is okay, but scaling while maintaining consistency is a nightmare. The solution here isn’t “the best prompt,” but scripts and validation layered on top of the model. And the stricter the technical requirements, the more work the scripts do and the less the AI has to do. A practical takeaway for those thinking about AI in development: code generation already works if you have tests and documentation. Art generation only works in a pipeline where AI is one of the stages, not the entire chain. And the percentage of manual oversight increases in proportion to how strict your technical constraints are. The game is available at this link: Code and resources: Source: [All articles in a row / Artificial Intelligence / Habr](https://habr.com/ru/articles/1065796/?utm\_campaign=1065796\\&utm;\_source=habrahabr\\&utm;\_medium=rss)