I have mixed experiences with the "chat" portion of Cursor AI. The "composer" has been much, much better, albeit on greenfield projects. I haven't explored it too much with existing monoliths, but I suspect it might not fair as well.
I'm currently using it to build an iOS mobile game, using Swift. I've never coded in Swift, or used SpriteKit, but it's going surprising well. When it does produce invalid code, I just copy the error directly out of xCode into Cursor Composer and 9 times out of 10, it fixes the issue first try.
I've used it on different languages, but the rust it produces, even if it runs, is often full of bad decisions. Choosing improper data structures, cloning everything, etc. Curious what the quality is on the scale of "running swift" to "good swift".
The code somtimes runs, which is a large improvement over recent years, but generally doesn't do what it's supposed to do.
And this is using claude 3.5, which works dramatically better for me in a chat session where i give it exactly what i want it to look at.
I just haven't found a way to be productive with it yet.
I've even tried the usual tricks of forcing it to write comments about what it's attempting to do throughout the code, but even then, fails.