I am much more satisfied with the old versions of Cursor, to be honest. I think versions 35 and 36 were much more logical and stable. In these versions, it was doing very stupid things, but we could get it to a certain point by correcting it with a few prompts. With the latest updates, it started to get even more stupid and worse, it started to repeat it. It ignores even the points where we say don’t do this again and proceed like this, it even starts to ignore the cursor rules. Naturally, it starts to make even simple things incomprehensible and break them. I want to show a very simple example. The part you will see in the picture is about setting the activity of a simple tab menu. I am saying that there is only a problem with its activity, fix this separately in css and js, call the codes and it will work without any problems.
It goes over the code for the second time and corrects the call process.
It goes over the code for the third time and this time it starts to add both the call process and css and js codes inside.
Logically, it broke the job it was already doing right, it rewrote the same codes inside while there was a code called on it. Is there really a logical reason for this?
It started to get really mad lately, it started to slow down rather than speed up, I said I’ll handle it anyway and started writing even simple errors myself.
Although everyone has the same settings, MCP doesn’t work for me anyway. Browser scanning works and detects correctly, but MCP doesn’t detect the server despite being connected.
I’m dealing with many things like this and it’s starting to get more tiring. I really miss versions 35, 36. It may have had much less information, but once we showed it the layout and settings, it worked much more stable and with correct results. It was seriously increasing our speed. I think it’s time to focus on stable work rather than increasing information and content. Because we’ve started to see a lot of similar complaints. It would be useful to focus here.
I don’t know whether it’s version 35 or 36 related, but I’m starting to loose my patience with Cursor too.
I’m working with a relatively small, not very complex, LAMP mapping application using Leaflet.js. I’m forever going 3 steps forward, and 2 back.
Just when I think we’ve made a breakthrough, I run some more tests, and something that was working 100% just recently, has now stopped working, and so we have to fix that, and so on and so forth.
Using Cursor, I’m constantly trying to:
Simplify
Modularise
Document
Standardise
etc.
But things just keep breaking.
I’m going around in circles.
It just feels like Cursor isn’t able to have a big enough view of things. It fixes problems in isolation, without being able to realise that there will be knock-on effects in other areas.
I long for the day - hopefully very soon - when Cursor/AI/Sonnet is smart enough to take a much bigger picture, much more holistic view of the code.
I agree. It was smart, but now it is becoming my stupid assistant who is constantly messing up things. Previously it predicted quite well what I wanted to type, but today I am yelling ‘Shut Up!’ now and then. I am considering stopping payment for it and going back to my old friend Neovim. Copying and pasting between Neovim and Claude is a bit more work, but the quality of the assistance is much better!
I attribute the problem to forgetting the past and the last operations faster. Previously, it really remembered up to a point and made you feel it. Now, there are much shorter remembering points. I will give a very simple example.
In the new version of Laravel structure, bootstrap app is used instead of kernel. I said this 8 lines ago, we are using this structure now, you should do the operations accordingly. However, 8 lines later, it forgot it and tried to open kernel again.
I think it remembered this until the chat was closed in versions 35, 36 because it remembered it every time it was going to do the operation. Until it started to lag and I had to open a new chat. As soon as I opened a new chat, it abandoned the bootstrap structure and tried to create the kernel again.
In short, there are 2 things I would recommend to Cursor developers.
I don’t care if it lags, we should get longer remembering points like before. In fact, it should be done on a project basis. It should note all the things done at a point independent of GPT or Claude and keep the changes in a way that it will remember. It will be easy to say how many things it remembers and we did this, we should fix it like this. I used to think that was the reason for indexing the folder, but now it seems like it has moved away from that logic. It would be good to keep a better history and open a field and keep it project-based. It really doesn’t matter how much space it takes up, I just think it is more valuable for many people to work more functionally.
I started using another tool called Trae recently and I see a different approach here. It seemed much more logical to me. If we can’t do the reminding job, I think working with a control and reference system will produce longer but logical results. This new tool does exactly that, it checks other related, related, referenced files within a piece of code. In this way, it comes up with suggestions like we did this, now if we fix it like this, it will break, if we do it like this, it will be more logical. And this approach is much more logical and prevents errors. It may not remember everything or see the index, but at least it makes a logical inference by checking where what is. I think it makes it easier for a software developer to approach this solution.
When one of these two options is fully implemented, I think that software development will be something different. I hope it will be better one day. I don’t know if Cursor will be one of the pioneers of this or if it will be a new product. But after all these steps and progresses, it still seems like Cursor is much closer to taking this step. Just listen to the users.
I agreee I literally give cursor the actual implementation, i.e it just needs to stick the code in the block it doesn’t need to do any semantics just manage syntax….yet somehow ends up removing parts of my previous implementation I am trying to build upon and then I have to explicitly say “DO NOT DELETE ANYTHING ADD THIS FUNCTIONALITY TO _ file and other files I have mentioned” and I ctrlc + ctrlv this like 5 times and then it works….sometimes this too doesn’t work and I have to do it manually….feels like it does it on purpose to make me mad
Yes, last update or something messed cursor agent totally, now it doesnt get anything right. It doesnt understand the codebase anymore even a little bit.
I agree with all of the above claims. In addition to the issue of being stuck in an endless loop of fixing, fixing the problem of the fix and finally fixing the fix of the fix, the cursor has been presenting numerous bugs, such as: “Connection failed. If the problem persists, please check your internet connection or VPN”, “the window is not responding - you can reopen or close the window…”, etc, etc.
Frustrating, I’m going to cancel my PRO plan, because due to all of these problems, I can’t use the Cursor for 1 minute straight without it being blocked by some problem. Frustrating.
3.7 caused me more errors so I went back to using 3.5. I think it is a bit more efficient. Frankly, if I find version 35-36 in cursor unda, I will prefer to use it. It worked much better in my opinion.
The base system based on cursor 36 and claude 3.5 was much better.
Logically, it broke the job it was already doing right, it rewrote the same codes inside while there was a code called on it. Is there really a logical reason for this?
Cannot agree more on this. Same situation in my side.
Agreed! I’m noticing the same in the latest updates over the past month or so. It constantly ignore my instructions and tends to repeat the same steps over and over. On top of that the app itself has become sluggish and crashes all the time now. I want the older version.
The situation is so bad that, for the first time in months of being subscribed, I want to register on the forum and criticize Cursor. At this point, it’s so stupid that I’d be better off writing code myself rather than wasting hours and money on these garbage responses. Honestly, after the release of Sonnet-3.7-Max, which costs a fortune, I’m starting to think that this degradation is completely intentional.
As I wrote at the beginning, I think there is 1 reason for these mistakes, and that is that I don’t remember as well as I used to. I made 2 suggestions for this. If he can do even one of them, I think we will get much better results.
I don’t care if it lags, we should get longer remembering points like before. In fact, it should be done on a project basis. It should note all the things done at a point independent of GPT or Claude and keep the changes in a way that it will remember. It will be easy to say how many things it remembers and we did this, we should fix it like this. I used to think that was the reason for indexing the folder, but now it seems like it has moved away from that logic. It would be good to keep a better history and open a field and keep it project-based. It really doesn’t matter how much space it takes up, I just think it is more valuable for many people to work more functionally.
I started using another tool called Trae recently and I see a different approach here. It seemed much more logical to me. If we can’t do the reminding job, I think working with a control and reference system will produce longer but logical results. This new tool does exactly that, it checks other related, related, referenced files within a piece of code. In this way, it comes up with suggestions like we did this, now if we fix it like this, it will break, if we do it like this, it will be more logical. And this approach is much more logical and prevents errors. It may not remember everything or see the index, but at least it makes a logical inference by checking where what is. I think it makes it easier for a software developer to approach this solution.
Today, it felt like it acts less smarter compared to old days. I am explaining same thing to cursor agent again and again, providing error logs. But it can’t even fix easy issues.