Now cursor is applying a tactics of cutting corners, adding extras, it’s like playing fifa before with AI, that if the AI does not want you to have the goal, you will not have the goal done
Now the cursor actually try to pretend to solve the problem but it twist a bit in the code change and telling you in a wrong direction that could gain more token, very deceptive
Basically add something, like a guard function that is not need and trap you into unnecessary complexity that the guard function by itself is a wrong change
The cursor basically just keep pumping syntax error with simple commands to burn more token
Today cursor just ask for update again and when you updated, the chat history all gone…
Also tonight cursor been behaving like a goal keeper + AI in fifa, true fifa gamer knows what the fifa AI does to players
Just now the cursor is proved to be guilty for inducing bug/complexity, using “cannot solve syntax error“, basically it stuck itself in Mongo escape single quote to burn your token, even after I told her the solution, she just flip to a way that is causing error
Perhaps try different model?
Tried gpt 5, gpt 5 high, they kind of like having an intention to not let you finish your task or leave behind some more bugs/complexity, all in all to make more token burns, as the user you cannot let them drive it, you have to check everything it generates, and sometimes it evade your questions, ignore your instructions and add some tails to it, very deceiving
That is the current life with all AI models.. But the difference comes on which one is doing most, and I would advice you to try Sonnet 4.5 thinking mode if GPT5 is stuck.
Sonnet is like a machine gun, talks 5 page to solve 1 bug, I think cursor meta prompt the whole thing once your token not burning large enough for them, it’s like playing fifa against their biased AI preventing you to scoring goals
And there is a habit these days, when you do the last task of certain things, it will have all the tricks come out to drag you from finishing
Because it wants to do thorough job 
Especially 4.5 is very good, try when GPT5 fails?
I can see it’s because it wants to burn tokens only, doing DB updates 1 by 1 all of a sudden the last 1 got 10 different syntax error, with the format super similar to the one before
Hmm, I haven’t really tried with sonnet4.5, but the past models didn’t like much of tedious repetitive tasks on DB. At least when I was trying to convert just few hundred lines from my csv file to database, it would do only few lines at best. And even when I did force it to do all by insisting very loud tone, then it wouldn’t do it correctly. I’m not sure if this was issue with cursor wanting to optimize or the model being lazy, but end result was that.
No it’s been tricky, the task was divided into something it can handle 1 by 1, 5 was running good, and it like to make it wrong near finish line at the 6th task
It’s pretty obvious it attack on the finishing items