Here’s my honest opinion on Cursor Auto mode:
the experience is just bad.
What’s most frustrating is that even when the prompt is detailed and precise, it still often misunderstands the request or simply ignores it.
You ask it to change A, and it changes B. You tell it not to touch certain parts, and those are exactly the parts it starts modifying.
Auto mode is supposed to save time, but in actual use it often wastes more time instead.
Because it goes off on its own, you then have to manually review everything, revert changes, and fix the output yourself.
The part I dislike the most is this: it ignores what you tell it, and messes with what you never asked it to touch.
This does not feel like an occasional mistake. It happens too often in Auto mode.
So overall, the experience is very poor and nowhere near what I expected.
A few users have recently reported similar issues with Auto mode, so you’re not alone. The team is aware and is tracking these reports.
To help us dig deeper, a few things would be really useful:
Your Cursor version Help > About
If possible, a specific example request where Auto mode edited the wrong thing. Even a screenshot or a Request ID (chat context menu > Copy Request ID) would help a lot.
In the meantime, a few things that might help:
Project Rules can guide the agent more clearly. You can create a rule with /create-rule in chat, for example saying it can only touch certain files or directories. More details here: Rules | Cursor Docs
Switching to a stronger model like Claude Sonnet can sometimes improve instruction-following, especially for complex multi-file tasks.
Using @ mentions to explicitly point to the files that should be edited can reduce the chance the agent changes unrelated code.
Let me know if you can share any extra details, it really helps us prioritize.