Struggling with AI Coding
I wouldn't consider myself a proper software engineer, but I've been coding for around 10 years, with the things I am coding increasing in importance to the business at the same time that AI coding is increasing in prevalence.
Even with a well-designed spec, I find that the code that AI generates is just ... I don't know. It's perfect when zoomed in, but as soon as parts start interacting and you are not coding from scratch every time, the whole thing falls apart. I mean, it will compile and work, but what comes out of it is riddled with incomprehensible layers of abstraction and errors. When I go in and try to understand those errors or explain why my code made them to others, it takes me days. When I ask the agent that wrote the code to explain what the code does, it seems to struggle even more than me. Fixing things without thoroughly understanding what the AI wrote (in detail), tends to make things worse.
I have tried modifying my .md files to include good software hygiene, agent review loops, heuristics .... none of it seems to make much difference.
Does the entropy that agents inject into a codebase eventually solve itself, and maybe I need to release my grip on the reins for a longer period of time? Is it just highlighting my own incompetence as a software developer? Or am I doing something otherwise wrong? Maybe I am just projecting my own issues onto the technology.
I am thinking about switching to something like cursor tab so I have tighter control over what comes out.
I don't see things going back to a time w/o AI assisted coding.
What has worked for everyone here?
Please feel free to roast me too :)