I have been programming for about 10 years. At first, I was not satisfied with the results, but now, after refining the process, I can instruct the AI to create individual elements, and it does so without my supervision.
When the component is ready, I only make minor corrections, and it is complete.
I’ve even gotten to the point where the AI codes in my specific coding style, just the way I like it. It’s all a matter of process and understanding the advantages and limitations of LLMs.
Why is this incoherent? It “is” broken doesn’t mean it was “always” broken. I honestly found the same thing recently, but only when using “Auto”. In fact, I only trust Cursor with Claude Sonnet 4 (and I’ve actually stopped using it almost entirely now, in favour of Claude Code).
My projects also got mashed by Cursor (in Auto mode)—I think it was GPT-5, but I can’t be sure. The problems seem to be related to projects that are a long way into development when the model starts with them. I’ve heard it’s okay with new projects. In my case the most infuriating thing was that it started adding really brainless fallbacks everywhere, making debugging extremely difficult. So it prioritized never crashing over correct execution. Super bad habit, imho.
He never said it ever worked. Even after being asked multiple times if that is what he meant, but he never said Cursor worked, it only messed up his projects. So he implies it was always broken for him and yet he tried it with 50 projects just to make sure it is definitely broken. It’s a bit illogical. It would make more sense that Cursor worked, so he kept using it, but then it’s capabilities got worse. But that was never said, so we are left confused.
You can see this user is basically a troll who stirs things up, says nonsense, and avoids clear civil discussions.
Ah, true enough. Sorry about that; I just saw the initial statement and didn’t go through the whole thread. But I can certainly sympathize with the experience of Cursor messing up a project. I find I have to be very careful about checking its work.
You never bothered to read my responses. I answered all of your questions already. I stand by my original post 100%. If anything I understated the issues. If you would like to actually address any of the issues I presented, none of which are at all new to the forum, then you ciould perhaps be useful/
I have read all your posts. You have not once said if Cursor ever worked for you and why you used it on 50 projects, if you already knew it was not working. Whatever. Something is simply wrong with you.
More times where this user starts threads and doesn’t answer questions when people converse:
I couldn’t agree more. I’ve been in software development for over 15 years, and I’ve never encountered a tool that causes this level of frustration. At first, Cursor performs reasonably well, but then the reliability just collapses — edits go rogue, files get mangled, and suddenly you’re fighting the system instead of building with it.
If this were a developer on my team, they would’ve been shown the door a long time ago. It genuinely feels like the product is structured to keep users paying while delivering less over time. I’ve reached the same conclusion — it’s time to cut my losses and move on.