The right and the wrong way to use AI [by codemonkey]

Codemonkey talked about vibecoding in his last Game Dev Report and I think this is an excellent example to be at least careful when using AI. It is not "DO NOT EVER", it is "USE WITH CARE". I really recommend everybody in here to read it.

One developer shared how they spent about 40 hours over a couple of months trying to "vibe code" a metroidvania in Godot, hoping AI could handle the technical side while they acted as the creative director. At first it looked promising, they got some basic movement and systems working quickly, AI nowadays is absolutely excellent at making good small code snippets, but once the project started getting more complex and systems needed to be connected with one another then everything began falling apart.

The AI struggled with enemies, combat, refactors, and broader architecture, and every fix seemed to create more problems. Eventually they gave up and concluded that full-project vibe coding is basically a myth, at least for something as complex as a real game.

And yes, that is what I am constantly saying. AI is excellent if you use it as a tutor to help you understand something. If you ask it to explain some piece of code line by line so you can understand then that's awesome!

But the problem is when people ask it to generate something and just blindly copy paste code. If you do that then you don't understand what the code is doing. And if you don't understand it then when something breaks (inevitable) you will not know how to fix it. If you're making something simple like Flappy Bird then it's possible you can build that whole game using AI without anything breaking, but for making any game even slightly more complex than that then it will break guaranteed.

When you're vibe coding you're not actually building the game, you are just stacking mystery boxes on top of each other until the whole thing collapses.

The developer asked the AI to generate a bunch of stuff and it successfully generated a working character controller that could run, jump, climb ladders, read tablets, use a scanner and more. All that done in just a few hours, impressive. But then when they ran into issues they tried solving them with more AI which didn't work, for example trying to refactor some code into a state machine just made a massive mess of global data. Since the developer did not understand the code they had no other choice but keep asking the AI to fix things and hope that it worked, and over time it worked less and less.

So the takeaway here is not "never use AI." It is: AI is not a replacement for understanding. Use it to brainstorm, speed up small tasks, maybe help debug something specific, but do not expect it to magically make up for not knowing the engine, the language, or the fundamentals of programming. If you want to build a real game, eventually you need real understanding, so use AI to help you understand.

1 comment