I’ve been using the previous version for a few months, and it was mind-blowing. For the last 10 years, I’ve been either CEO or CTO and haven’t been coding much. Everything changed with Cursor
I’ve been coding for 10 hours straight every day and enjoying it like never before.
However, a few days ago I upgraded to the latest version (I haven’t read any feedback before that) and I’m struggling every step of the way to my MVP getting ready.
In the new version, when I asked the composer to add a new table and update prisma files, routers, controllers, etc, it did not follow the existing file structure, it wrote code in the wrong files and gave bad advice all the time. Still tremendously better than me, but the quality drop is evident.
I forgot to mention, the tab code suggestions were unbelievable. I felt it was reading my mind. You didn’t have to type almost anything, it knew exactly what needed to go into the next line. Brilliant! In the new version that completely stopped. Sometimes it gives some suggestions, but they don’t follow the existing code style and naming convention and most of the time do nothing.
It’s even worse when I work on my Chrome extension. Anyway, my point is not to shit on the game-changing product like Cursor is, just give feedback on it so you can make it better.
Agreed, it’s so weird but 0.43 provides a lot worse suggestions and forgets what we’re talking about in the same chat instance, only after 1-2 suggestions that don’t work, it starts suggesting new ideas that are completely off topic.
I’m a huge cursor fan, I hope they can fix it because it seems to have fallen down to “regular” AI tooling capabilities which are cumbersome. I find myself actually turning to Perplexity directly for a lot of questions now because cursor is starting to put out unintelligent responses. It used to be so smart and so predictive… maybe it was too good to be true?
I hope they make it possible to downgrade to 0.42 and stay there. Because of the auto upgrade hassle, I am using 0.43 and it’s making me fall out of love with cursor.
I totally get how you feel, I’m in the same boat. 0.42 was amazing, I even bought a one-year subscription right away, but just a few days later, 0.43 came out, and I’ve been facing some issues. As a supporter, I’m not asking for a refund, but I really hope the developers can come up with a fix.
I’ve gone back to GitHub Copilot today and plan to try out other alternatives over the weekend. If you have any recommendations, let me know!
no matter what model I select now… it is like talking to a ten year old. tasks that claude sonnet previously handled with ease (particularly when dealing with a codebase/long context chat) - now it just asks loads of dumb follow up questions - “would you like me to proceed?” - then does half the task and asks more dumb questions with “would you like to continue” - it’s awful and to the point i might even just cancel the subscription. what’s the point of it now? I can get this level of stupidness for free.
Having reverted to the previous version 0.42.5 - and submitted the EXACT same question, the EXACT same files and using the “long context” option with the claude sonnet 200k model - i was able to get the results i expected.
To be clear… I have a 4000 line powershell module. I am lazy and don’t like writing README documents. I get the AI to do it for me. Using the CHAT/Codebase Chat in 0.43.5 with Claude sonnet selected… it will continually ask follow up questions, perform a part of the task and ask more dumb questions and generally acts like a wayward preteen who’s been told to clean his room and is pretending to do it in a half assed way. I do the exact same thing in 0.42.5 with long context and off it goes… generates a really concise README with ToC, usage examples… output examples…
They managed to mess it up. Bring back Long Context.
The previous version was incredible, but since upgrading, it feels like it’s not using Claude 3.5 Sonner 20241022 model anymore, because the code output and understanding is really bad compared to before.
I also have Claude’s own License and compared same query on same model through Cursor (using Claude) and directly against Claude. Claude understood immediately and gave me a very high quality response that worked, while Cursor (using Claude same model) could not even get to a working solution by me giving it several hints on what it should do.
This worked before on Cursor (using Claude) perfectly
I have bought licenses for all our devs and now it’s basically pointless and risking bad code quality and bugs.
I don’t think Cursor is using the correct model, even though it says it is.
Additionally I want to mention that I feel that the Composer also has become slower now when it’s not the floating window anymore. Before it was writing the code immediately into the code, now it feels like it spends the time writing it into the chat first and then additional time writing it into the code.
Just feels slower, but I could deal with the slower version as long as it is the quality of regular Claude Sonnet 3.5 20241022 version.
I have also ensured that I have fresh account with plenty of premium requests left
Seems like they’re dropping context more often and more quickly. I have to reselect files in chat multiple times to keep them in context. Cheaper for them, but much worse UX, less utility in the product in addition to worse code gen
Question: I was having issues getting the README.md to show up properly in the chat. I think that Cursor uses markdown to display the chat, so the README that was getting generated was displaying instead of being listed as pure code? My workaround is I made a README.html file instead, and then I’ve been using pandoc to convert to markdown. its clunky. how are you generating READMEs?
and yeah I’ve noticed a sharp decline in context ability int he last couple days. Even when select the file in the little drop down of the context window of the chat, I still have to copy paste the file into the chat window. it wasn’t like that a few days ago.
Yes thats what exactly is happening . I think they have drastically reduced the context because of the agents. so frustrating as the agent easily looses context
Generating readme can be a bit hit and miss. I’ve found if I ask it to generate markdown it will attempt to comply and then in places where it attempts to use triple backticks the message appears to break out of the code block… all I do is let it finish and then copy the entire message and paste into my readme - you’ll see the initial '```markdown" - which will need deleting, along with anything it sent prior to that… from there t is just a matter of cleaning up the formatting of the broken code blocks that copied without formatting. It’s not perfect but it works.
I’ve also found that passing 4 or 5 functions into the chat at a time works a lot better than asking it to document an entire module, for example.
My advice to everyone is to simply roll back to 0.42.5 because this newest version is awful. The difference is like night and day. As soon as the 0.43.5 update hit my machine I knew something was off and it only took me a day of using it and I made the decision to roll back to the previous version.
Cursor if you’re actually reading this you made a huge mistake and using the chat in cursor 0.43.5 is not something I would continue to pay for. If I want that kind of AI assist I can pop over to chat.com and get it for free. If there wasn’t a way to roll back to the previous version I would’ve cancelled the monthly and found an alternative… Then I’d be here complaining bitterly about how crap it was and would be encouraging others to ditch it too… We pay for this. Remember that… if it’s not fit for purpose or downgrades… we can stop paying for it.
I pay for the convenience and speed at which I can work… if it’s slowing me down by being dumb I’m better off without it.
I really hope the cursor folks are paying attention to this
Totally agree on this post. I’m not sure when they introduced auto-update, but Cursor auto-updated and is now missing long context features and the code generation is significantly worse. For example, we were able to generate an entire Mermaid diagram from a database schema that was thousands of lines long in a previous version of Cursor, but this new version struggles to get through the first 100 lines now.
I’ve been willing to spend some decent cash on this product, but this is a big step backwards and will reduce my hard-limit of usage-based pricing
This worked before on Cursor (using Claude) perfectly 
