There are more and I will post them as I rediscover them again.
completely agree.
Bugs with the remote VS code as well.
Had to revet to VS Code as Cursor is unusable at the moment.
However, Grok 2 is running very nicely even in Agent mode. I didn’t think it was supposed to run in agent mode.
I updated from 0.46.5 to 0.46.6 and now claude 3.7 is actually stupid. it messes up on simple tasks.
When Sonnet 3.7 stops thinking, it’s like 4o-mini mode is on without any notification. When asked who it is, it still confidently says it’s Claude and deduct fast requests credits. Just ask a simple question like “python quicksort” and compare it to the answer from model 4o-mini, they’re exactly the same. Cursor is blatantly scamming, refund please!
Yes, claude sonnet 3.7 is OVER ENGINEERING EVERYTHING. Unusable
I not using this forum very often.
I agree that the cursor 0.46.x is good but with Claude 3.7 together useless.
I disabled 3.7
Have somebody ever had some feedback from people at cursor?
An admin or something like that?
I’ve had really good results with the update and using 3.7 but it does work best with a persona prompt that’s specific to the work you are doing. Make sure the persona prompt runs an audit of the files you will be working on then create a plan and add temporary documentation to refer to throughout the chat.
However, it keeps stalling out on a single linter error. Has anyone figured out a workaround?
What’s worked best for me was reverting to older version of cursor 0.45x and sticking with claude 3.5 sonnet for now.
and they have been sending updates every single moment.
where did you find the older version ?
i just uninstalled the update thru windows as it showed up as a separate install, but you can find older version I heard here: GitHub - oslook/cursor-ai-downloads: All Cursor AI's official download links for both the latest and older versions, making it easy for you to update, downgrade, and choose any version. 🚀
Always check to download links containing official cursor.sh server to prevent hijacks…
About 3.7: it feels like agentic reasoning, you’re getting bad results like overengineering because your prompts are generic, check livebench.ai and see the difference of IF(instruction following) between 3.5 and 3.7, o3-mini-high is still a little better in coding/IF with a big difference, Claude is unbeatable for agentic tooling and that’s why its the best for use.
dont even need that github repo if you only want the previous version. Previous version is hosted on cursors site.
I am not having this issue at all
I think you’re experiencing reactance where you are used to an old thing and instantly dislike the new thing, like how people hated “New Coke” when it was the exact same formulation
I think what we liked about composer is that it seemes to make effort to understand the entire codebase. Like a composer in an orchestra, it would guide all the different instrument players, whereas the agent just plays one instrument really well
We all want AI to do as much of the hard stuff for us as possible, and we know it’s smart enough and capable enough to do so because we’ve seen Cursor at its best.
The people feeling cheated and asking for refunds aren’t scammers or bad people, they are simply attached to the projects they’re building and don’t like “feeling” less able to achieve them, even if it is a feeling.
It’s called Baby Duck Syndrome, and it’s one of the biggest barriers to UX innovation.
Maybe that is a point, but maybe the truth lies between. The thing I came across with 3.7 is, that it loves to do things I did not say it should do. It bloatens the code with functions I never told it to add.
That would be kind of fine and a new work-style IF the implementations would work. But they don’t and it just bloatens the code more when it trys to fix it, without fixing the actual problem.
But maybe I just did not get the trick yet, idk.
It is mostly doing things I did not explicitly ask for and that is frustrating because Claude 3.5 worked perfectly and should continue to work perfectly even with an updated version of Cursor and the addition of 3.7.
For example: I have a php composer file and view. I reference both of them and tell Claude to create a new composer and view based on the files that already exist but change some basic data and explicitly tell it not to modify anything related to the structure or design of the file.
Instead it does the exact opposite and creates a completely different page layout.
I then yell at Claude: “Why did you not listen? Why did you completely redesign the page when I specifically said not to?”
Claude responds: “You’re absolutely right! I should have listened better the first time and created an exact copy of the page without modifying the layout or design. Let me fix that for you.”
Then it does it again, mostly correct.
It’s frustrating because I am mostly happy paying the $150 to $200 per month in premium requests to have it spit out basic tasks like the one above, which helps me move along through mundane tasks quickly. For harder tasks, I will just complete them myself or get some help from “Ask” in chat. Gosh.