Comment by jongjong

I love using Claude Code. What happens is if your codebase is bad, Claude will produce bad code which follows the same general approach.

So if the foundation is overengineered, it will produce over-engineered code on top. If your code is super clean and succinct, it will produce code that is equally neat and succinct.

I've used Claude code with my own backend serverless platform and it's like it can read my mind. It's almost exactly how I would do it. It's like working with a clone of myself. I can implement nuanced features and ask for very specific changes and the success rate is extremely high.

On the other hand, I've worked with some other people's code which was slightly over-engineered and it struggled. I managed to achieve what I wanted with a lot of manual intervention and careful prompting but it was a very different experience and the code it produced didn't look like my code. It looked like it was written by a mid-level dev.

I can sense that it's a code completion engine because it clearly doesn't have an opinion about code quality. It doesn't know what metrics to aim for. It doesn't aim for succinctness; just as most human developers don't aim for it either.

My assessment of LLMs is that they don't have a specific intrinsic goal so the goal/motivation is inferred from the text it sees (from the input). If the original author of the code was motivated by creating complexity for the purpose of lock-in factor and career advancement, the LLM will produce code which optimizes for that. Unfortunately, this describes the vast majoy of code which is produced.

You can tell a lot about someone based on their code.