One big issue I have with using Gemini to help me code is that it's not very good as giving me the code in the manner I want. Either it outputs the entire script again making changes it shouldn't, or it outputs confusing snippets without clear direction as to what to do with it. So, over time things that worked may stop working or start malfunctioning. 

One example of this is the pause menu (WHY DO I KEEP TYPING "manu"?!?!) I implemented yesterday. It was working just fine when I opened it, and when I used the buttons, but hitting escape to close it wasn't working. So, what did Gemini suggest? Changing a bunch of the gdscript in the function to account for something that didn't actually matter.

But what was the source of the issue? The stupid AI decided to get rid of line that set the process_mode to always run on my autoloading game manager singleton at some point. So this process that was meant to track all sorts of things just decides to pause itself because it can. Kind of reminds me of back when I was dealing with my background music and the AI ripped out a section that fixed a problem I was having while we worked on something else and caused the problem to come back again.

It's really important to know what you're doing when coding with an AI and to very carefully look over every single change they suggest. I try to do this. I compare the original to the new, line by line, but it can be hard to validate everything because of all the extra crap it does. It's even worse that even when I tell it to forget previous versions of the code it still remembers and references them. And it's even worse when it starts believing (after being told otherwise) that it's suggestions have already been implemented.

But, that's what I get for choosing this approach.