[alexpotato]

I've started trying to make some vibe coded games.

The best analogy to use is AI image generation e.g.

- someone posts a photo made with AI and your reaction is "that's amazing!"

- then you generate an image and it's pretty good but not exactly what you want

- then you try to get the AI to change the image to what you want and it's very difficult/requires a lot of steps

More specifically, you can easily spin up a game and parts of it are great. e.g. I made a game where you run a Soviet tractor factory [0]. The AI came up with some good mechanics and funny scenarios ("inspector comes to visit from Moscow: bribe him" etc). But the game mechanics are "off", the humor doesn't always work, you need to gametest A LOT etc.

0 - https://alexpotato.com/games/tractor47/?l=hn2

[contextfree]

I've been experimenting with vibecoding with a 3D game with no engine, all in C++ with DX11, all-procedural modeling. My experience is you can quickly add a lot of features and mechanics that work semi-well but getting any one feature or scenario to work really well ends up "requiring" digging into the code and writing stuff yourself.

"Requiring" in the sense that while it might be technically possible to get the AI to do it, it becomes an incredibly annoying experience to repeatedly try and fail to get it to properly understand and implement what you want. This almost seems less about whether the AI is as smart or smarter than you or a better programmer or whatever, and more about the medium of natural language being worse than the medium of code for this.