Yesterday, I burned $25 on Claude Code trying to debug a long-standing issue I had in a project. It started impressively by understanding and explaining the code base better than I even could have. It had a firm grasp of everything, even small details and obscure options. Then I let it refactor my code and everything fell apart. The worst part was that after every time trying to fix the issue, it confidently reported success, only to admit that the issue was still there when I pointed it to it.
It couldn’t really solve it. In the end I wasted four hours of my time and $25. It was quite a let down.
The code it had written was ok. It wasn’t fancy, but not horrible either. About advanced junior level, I’d say. But the code didn’t really do anything, it was a lot of smoke and mirrors.
Don’t let anyone fool you into thinking that LLMs will write software on their own soon.