I fully agree. It is particularly true for the agent mode in which cursor (I use Sonney 3.7) insists on rewriting hundreds of line even when a tiny change is required. I need to specify: reload files A B C to parse the modifications I made, and only change things that are entirely required for the task at hand.
Weirdly, if I put this in the rules, it gets ignored, so I have to add it every single time, which drives me nuts, and half of my requests end up with a rollback.
I have that issue only with 3.7 and its a well known issue of 3.7 no matter if in Cursor or elsewhere. Thats why i use 3.5 which works without this issue.
1 Like
Recently I found the following solution, continue using 3.5. Also, be sure to select the folder or files to be changed. For example, if the change is index.html, style.css, script.js, add these 3 files and then when you say I want to fix these parts in this file, it offers more logical solutions. Thanks to this, I was able to solve a little more problems again, but it does not index our entire folder like before (even though it says it does) and the past does not remember well. I hope the old functionality comes back. We were much more productive in the old way.
Honestly, a few versions ago Cursor felt like a miracle. Now, it doesnāt even seem capable of scanning my codebase on its own. When I give it a prompt like:
@context @context Do X Y Z
The response is typically: āOkay, this part does this, and that part does that, butā¦ hmm, I donāt understand how this component works, so I canāt complete the task until I do.ā
Previously, it used to work like this:
āOkay, first I need to understand how that component works.ā
[It would then scan the relevant code]
āOkay, now I understand. Letās do X, Y, and Z.ā
Even when you manually provide all the necessary context now, it often just produces a complete mess.
What really frustrates me is the complete silence from the Cursor team. No one is addressing this. Thereās no āWeāre looking into itā or āWe had to nerf performance because XYZāāabsolutely nothing. Because of this, Iām cancelling my subscription next month.
2 Likes
I agree with the community. When I first started using the Application about 2 months ago, it was a miracle app and model. The language model felt intuitive, smart, understanding, had great recall and followed instructions very well building apps left and right (this is not Cursors fault). The app hardly ever slowed down, hardly crashed, but now it can barely start up on my projectsā¦ after each update something else get slowed down (Cursors fault)ā¦ its definitely slowing progress thatās for sureā¦
1 Like
Cursor has turned into a very unproductive game of whack-a-mole!
I moved from Bolt to Cursor in late November and was thrilled. After a brief defection to Windsurf, I came back in late January. It was a delight, and very very productive.
Since 3.5 however things have gotten really out of control.
My experience is that I wall be working on one specific issue and Cursor will alter (and frankly devastate) other parts of the code, wiping out functions and re-writing the UI. The leads to seemingly endless testing, fixing, breaking, finally working, other stuff unrelated breaking, fixingā¦and on and one.
I know these are still very early days for vibe coding, and cursor has been in the pole position, but the past four weeks of chaos is driving me back to windsurf and heck, maybe even bolt. I am really sad about this as I love the whole way cursor works, when it does. For now, it does not.
1 Like
Indeed. Cursor used to 1-2 shot everything I needed. Now Iām spending 10+ requests to do the simplest of things and still getting some sort of abomination of what I asked it to do. It constantly forgets what it was doing, just plain disregards instructions all the time, tries to redo what it removed in the previous step. Just useless.
I wonder if there is some sort of model spoofing in play - Cursor might tell us that we are using Claude 3.7 but since it is so expensive they are substituting it with some of their own crappy models, only to send the next request to the actual big model, aka the model auto-select but hidden and forced.
Regardless, personally Iām real close to moving to Windsurf as the Cursor experience at the moment is just abhorrent. And just plain infuriating as F.
2 Likes
I just didā¦moved back to windsurf and I am rocking again. Perhaps this is just the nature of vibe right now. When our tools destabilize, switch. But know that next month, the tides may turn again.
1 Like
URGENT - please, can anyone confirm that post someone said: Model Spoofing
āI wonder if there is some sort of model spoofing in play - Cursor might tell us that we are using Claude 3.7 but since it is so expensive they are substituting it with some of their own crappy models, only to send the next request to the actual big model, aka the model auto-select but hidden and forced.ā
my feeling is like, the model itself might be spoofing us, anyone can confirm???
1 Like
Yes, new crazy Cursor bugā¦
It gives the generated codes as code and expects me to make the changes. Great, I can do the same with Claude or Chatgpt. I thought the times of asking the question and copying and pasting were over? Why did we set up Cursor if I was going to copy and paste anyway?
4 Likes
I totally get the frustration with recent changes. Weāre seeing a lot of feedback about issues with context retention and unnecessary code rewrites
For now, here are some things that should help:
- Use Claude 3.5 instead of 3.7 (seems more stable for most users)
- Explicitly select relevant files before making changes (Cursor ā Welcome to Cursor)
- Regularly start new chats when working on a different feature, file or area of your code
- Use checkpoints frequently (Cursor ā Overview)
The team is actively working on improving context handling and preventing unnecessary rewrites. You can track known issues and updates at https://cursor.com/changelog
For the specific copy/paste issue - try using the āApply Changesā button or Ctrl/Cmd+Enter to apply code directly rather than manual copy/paste. More details at Cursor ā Apply
6 Likes
thakns Dan
One thing that could help would be to let us set the query limit to a lower value, so that it could be easier to control the output before we have 2000 new changes to review.
From my point (creating a SaaS platform with react, next, tailwind and supabase with APIās, itās getting worst and worst with this new update.
I had to stop my development, because it was starting to destroy everything we had created after almost six months. Let me explain and give you some points to not mess this up!:
1- sonnet 3.5 seems to be much more stable, iāve been using 3.7 MAX (thinking it will be better) but do not waste your time/money mate! Use 3.5 for the last Cursor version.
2- READ, READ, and READ. If you have any MCP, which allows Cursor to take control and walk around like Peter in his house, disable it, and above all, do not use the Yolo Mode, you can make a big mess. You ask him to do something, he runs it, you test it to see the errors and you tell him that he answers you, that he already knew it, I have passed it many times in the last 48 hours and it makes me really ā ā ā ā ā ā off knowing that he has the knowledge that something is going to happen, not even putting him the tests he is not able to do it by himself, until you ask him, it is as if he had suddenly lost the ability to react.
3- Select files and give ALL THE TIME A LOT OF context of what he is doingā¦
4-Checkpoints will save your live, also Github and branches, Gitflow and so, or you are really ā ā ā ā ād up!
4-NEVER USE MULTIPLE CHAT TABS OR CHAOS WILL COME TO YOUR LIFE!
5-Every time Cursor tell you to start a NEW CHATā¦ DO IT!!! or your soul (Project) will be taken by a junior programmer.
For the restā¦ last chance for Cursor before moving to another.
1 Like
I appreciate you sharing this and @danperks too.
I get there is frustration but this thread is awesome. Instead of complaints I just read some good advice.
FYI, I also ran Claude 3.5 last week with usual success over 3.7 which feels nerfed in the last two weeks??!!
1 Like
Unfortunately Cursor is not working well for me too and is having strange behaviour when trying to apply changes to files.
I noticed that now Claude 3.7 does not apply changes properly. It just re-reads parts of the same file. This is just a screenshot of it just reading a file and claiming it makes changes to the file. It goes around in circles.
The list of āRead index.jsā actions is way longer and could not fit into the screenshot.
When i said to it that āyour changes were not applied for the fileā it started reading the file again for a couple of times and then it wrote a diff which i had to apply manually.
Claude 3.7 feels way worse:
- before 3.7 it used to not read parts of files every time it needed to do some updates
- was generally feeling smarter - i am now worried that it will mess up after each prompt
- i have to spend time correcting it a lot, up to a point where it might be better if i just write the code and have it look over it (used to be vice-versa, it wrote the code and i did the review)
Anyway, iāll try switching back to 3.5 and turning off MCP - the forum has been very helpful in identifying common issues.
2 Likes
You can downgrade!! Iām using 0.47.4 and its working really well with claude 3.7 thinking. Just every time when I start up I have to unselect the Auto mode, not sure what it choses but itās bad..
You do need to disable auto update in VSC settings though (json), otherwise it updates and you start using the new version and are like, why is it suddenly so stupid, until you realize it upgradedā¦
Even with all kinds of rules in place, even with clear constraints and boundaries explicitly defined, it still frequently messes up the code, often turning correct code into complete chaosāutterly like a deranged developer, making the experience unbearable!
1 Like
Cursor doesnāt realize itās cursor anymore, even after using 3.5. It thinks itās copilot. It canāt see the whole IDE.
EDIT: I also have found that all functionality is restored with 0.47.9
2 Likes
lmao so true