Cursor results are going to depend heavily on the model; Gemini 2.5 pro exp seems the overall strongest. You’re probably defaulting to 3.7 sonnet which is completely unusable; it was good at first but I am convinced anthropic “updated” (degraded) it behind the scenes to lower their inference costs. OpenAI did the same with GPT-4o for a bit a while back before making it better again.

3.7 also seems to have converged more on the hybrid reddit user/npr listener/HR lady tone and manner of speaking that makes me want to punch a wall. Genuinely people could probably increase LLM usage just by fixing this problem and banning r*fit from the training set.