The agent seems to not get anything right.. even the simples code base it constantly gets things wrong.. reintroduces errors .. refuses to look up or take into account documentation I give it.. constantly changes my models to older more expensive ones. Completely forgets what we are doing if I stop working on it for only a day… is creating bloat code like never before.. I asked it to simply change a function and move it to a new tab with 3-4 UI elements being moved to.. it created 3400 new lines of code. it really feels that it is unusable. I will see how it plays out this week but if it does not get back to normal I am canceling my subscription and moving on.
10 Likes
Same. I’m considering removing the subscription for the whole team
9 Likes
Same. I asked it to move two methods and it has totally borked my project 6 times and I had to revert everything and start over.
2 Likes
Long story short, Cursor is losing a ton of money and had to make drastic changes to how the app works, resulting in drastically shorter context and output capability. As a result, we’re seeing all sorts of degradations across the board and various situations. Unfortunately, this isn’t going to get better - that’s why they haven’t fixed it. It’s a necessary cost cutting measure. The only solution is to move onto a more expensive solution that, while being pricier, will perform the way you’re used to Cursor performing in the past.
4 Likes
Completely opposite
Using it for a year and last few weeks with Sonet 3.7 was the best
Just today was able to hop on new repo that never worked before and fix a few bugs with no issues
I have tailored tuned cursor rules and I observe how it works and if it get carried away I’m correcting him before he’ll change 3000 lines etc.
1 Like
My problem was it didn’t stop when I clicked stop. It didn’t stop when I told it to stop and it even AKNOWLEDGED it should stop and didn’t 5 times. I had to close cursor and restore what it did.
2 Likes
Sometimes it overengineers stuff but its a problem by how the prompt was formulated, if its true what @co50 says about output capability it would not output 3k lines like what happened to @iisjasey, I think its an issue with how we’re prompting 3.7 as it needs more defined prompts, @iisjasey try the below prompt when you want to refactor some code and modify parts to better match your task:
You are a powerful model specialized in refactoring Python code. Code refac-
toring is the process of improving the internal structure, readability, and
maintainability of a software codebase without altering its external behavior
or functionality. You must output a refactored version of the code. Explain
the steps you took to refactor the code and why you selected the refactoring
type/types you did.
# Suggested refactoring types:
example: "Consider using Extract Method or Inline Method"
other types: Extract Method, Rename Method, Rename Attribute, Rename Parameter, Rename Variable, Extract Variable, Inline Method, Inline Variable, Add Parameter, Extract Class, Parameterize Variable
# unrefactored code:
@pythonfile.py
3 Likes
Thanks! I will try that prompt out! 
1 Like
Do yourself a favor and cancel. The team is not interested in helping.
5 Likes
I doubt this is mainly a Cursor problem. I get wonderful results and also it fails sometimes, but when it fails I always think its a problem with the model my prompt and the context I’m adding to the prompt.
Valid point—Cursor itself is rarely the root cause. More often than not, it’s a combination of the model, the prompt structure, and the context provided and 85% of the users either use 2 of the 3 mentioned above or none of them.
2 Likes
I agree, it’s really frustrating to see how it keeps forgetting instructions, cowboy code over a very carefully planned app structure and breaking stuff for no reason. It’s uncontrollable and very tedious to use, I’m faster coding by hand at this point.
Following post from another forum member likely applies here
2 Likes
I’ve had this problem, too.
He doesn’t listen to me.
I asked him to look at the project structure file, but he didn’t look at it. He chose to search by himself, wasting time.
I gave him files and he didn’t read them. Instead, he chose to read other useless files. I don’t know what the problem is. It’s very difficult to use now.
claude 3.7 sonnet
Try to use 3.5 instead of 3.7 if you are not sure how to use it for your project
What i notice is the whole change from removing the @codebase seems to have shifted it into a worse position for intelligent code production. Not sure what is happening under the covers now, but I find that it takes us so long now to enter all the relevant classes so it can intelligently make a change its becoming too burdensome. On top of this, even when providing a class contextually the model/cursor doesnt even both to review the class and makes up its own new class on its own even when one already exists and was provided it. All in all to say that we have blown through our calls with horrible results and spent all our calls on correcting errors it produced. This is not the cursor I have enjoyed for the past year.
I hope they see what has happened and do focus on improvement. Others are passing cursor by now and my team has to start evaluating alternatives.
cursor must stop allowing trial users to use the advanced paid model. A large number of unlimited trial users have drastically increased the API costs of cursor, so they are making the real paid pro users pay for it, reducing the input/output length and performance of the model, which is simply unreasonable. In that case, I think maybe I need to be an unlimited trial user.
1 Like
for me it’s like 70-80% of my request end up failing
so I waste around 400-500 request every day in stuff that is actually never done, leaving aside that most of the time I have to cancel the implementation of what it actually does… I get that if cursor is loosing money he has to do some cuts, but making us waste money is not the solution, its a loose loose situation
1 Like
Do you have any proof of what you are claiming? or are you just inventing fake reasons?
Here are some relevant reactions. First, a post about 10 days ago began to clearly feel the rapid degradation of the capabilities of 3.7, including 3.5. There are also a large number of feedback discussions on sub-problems in the top posts of the forum. Second, the cursor-auto-free project once appeared in the weekly rankings of github for a period of time, with a rapid increase in the number of stars. Users can use this project to easily achieve unlimited free use of advanced APIs of cursor. Pro subscription users only have 500 requests per month, but they have unlimited. Considering that cursor needs to pay anthropic for each call, which has caused a large increase in costs for cursor, it is reasonable to suspect that cursor has degraded the performance of the Claude API (such as reducing or compressing the context) in order to reduce costs. It is difficult to understand that they do not restrict the way to try the advanced API for free.
