It strikes me as a legitimate reason why its getting worse! That or reverse engineering, but surely Cursor wouldn’t do such a thing would they!??
I think complexity of all the things combined (prompt, model , rules, mcp, cursor tools, context, files etc…) can cause it.
Once i simplified some rules and requirements broken down better it worked again.
I just switched to 3.5 and its sharp as a tack!! hahaha. This thing is so touchy!!
1 Like
3.7 in general is a bit hallucinating too much and trying too hard to be helpful, even outside Cursor.
1 Like
Must have scraped its definition of helpful from a government website!! lmao
1 Like
3.5 is flying though ain’t it @T1000 ?!
its great, it ran so many commands in row without errors. Sure a few times had to guide it a bit back on track but nothing major.
the issue with 3.5 is that it tends to forget half the stuff, like if you give it 3 or 4 commands, it will forget half at least, not even talkiing about following rules.. yes it goes fast, but does whatever it wants, fixes one problem and ignore the rest,.. or the consecuences of the problem itself
is fun when it says.. "we are done! everything was implemented following the plan.. " and you have 492 errors on the file
1 Like
I suggest breaking the task down into smaller or clearer definitions. Ask it to follow SOLID and DRY principles and to not write more than 500 lines per file. Pretty much correct with 3.5 then