I assumed there would be a continous improvement process whereby what worked amazingly well would be retained and refined further, yet the day-to-day experience with Cursor is so different that sometimes it feels like im using different tools.
One day it would understand my prompts exactly and it finds the files without any problems and does it’s magic, then suddenly the next day it deletes large pieces of code or gets stuck searching for files, or finds a terrible solution to a simple problems etc etc/
ive alternated between claude 3.7 and Gemini 2.5 pro.
The only one that i have had slightly better consistency with is 3.7 [max] but i cannot afford to keep using that
Cursor performs brilliantly with Claude 3.7 - with the tradeoff being cost. Hopefully cost will come down significantly as has been the trend in the past.
Claude is intended for analyzing complex tasks - and it does that well.
A workaround for now may be to observe in more detail which tasks can be performed consistently by less expensive models vs. which ones require 3.7.
Another workaround is to break your queries into smaller tasks which a less expensive model is able to. perform well.
Moreover if Cursor is behaving inconsistently consider setting up a README file and/or a Cursor Rules file. If Cursor does something wrong or inefficiently prompt cursor to update the README file with that example so it does not repeat the error. And set up a rule for Cursor to review the README file each time a new context window is opened.