Hey folks,
Over the past couple of weeks I’ve noticed that even with carefully engineered prompts, most LLMs are spitting out broken or off-target code. I’m using Cursor with Golang, Python and TypeScript projects and am hitting at least one of these on a daily basis:
- Deleted blocks that were never touched by the request
- Logic errors that unit tests catch immediately
It feels like the underlying models have regressed or are being switched behind the scenes. I’ve tried:
- Switching between OpenAI, Google and Claude
- Adding explicit system messages about project structure
…but the hit-rate is still low.
Questions for the community
- Are you seeing the same degradation? If so, when did it start for you?
- Any reliable work-arounds (agent mode, chunked prompts, older checkpoints, etc.) that actually move the needle?
- Have you found a particular config that still produces solid code?
It just good for creating web templates in these days 
Would love to pool observations and maybe escalate a concrete bug report to the Cursor team.
Thanks!
Reza