I would say the problem is twofold:

- current technological limitations of LLMs

- how industry uses LLMs

we should not be working with code at all, the new level of abstraction are requirements and conversations with AI, this is what should be commited into git and then you would simply "compile" your conversation and requirements into code

and if you would focus on requirements and prompts as the new "code" maybe you would feel like you are getting some power and understanding back?

but obviously because of many reasons such as context windows and hallucinations this is not possible... yet