The new version of Cursor is not good... and Claude 3.7 not good

Ideas Feedback
22

agreed, always crash, terminated, stopped and no response.

23

Also:

I use Claude for a lot of basic front-end work. I will tell Claude to specifically modify one section by referencing code comments for sections and it will sometimes do what I ask it to do but also not what I asked for and will respond with something like:

It looks like the edit was applied incorrectly and replaced the product grid section instead of just updating the tabs. Let me fix this by restoring the product grid and properly implementing the new tab design in the Japanese Craftsmanship section.

Then it will proceed to think and attempt to fix itself but never does and never stops thinking.

I have to often restart from my initial prompt several times before it gets it right or end up fixing it myself.

1 Like
24

upgraded to 0.46.7 and it seems solid to me now. havenā€™t experienced issues just better experience.

25

Just realized they pushed an update. I will give it a shot, thanks!

26

Still not great. It just does not do what I tell it to do. I say, read this documentation. Instead, is just skips that and hallucinates who it thinks it works. Not great. I think their tool calling infra is just not well tested and tuned.

1 Like
27

I experimented a bit between 3.7 and 3.7-thinkig.

3.7 in agent mode gives often better results on simple short prompts.

For 3.7-thinking you need a different type of prompting because of itā€™s reasoning nature.
sonnet 3.7 Normal Prompting: Prompt engineering overview - Anthropic

sonnet 3.7-thinking Prompting: Extended thinking tips - Anthropic

I hope this helps.
In combination with mcp and the docs feature 3.7-thinking in agent mode is able to create nearly complete packages for neos cms and neos flow. What is a huge task for most models. :flushed:

3 Likes
28

Applying each file separately has become necessary. Apply all button is not appearing.

2 Likes
29

What docs feature are you referring to?

30

I completely agree with you, but a lot of this can be negated with a good rules file.

31

This +1. New behaviour is stupid

1 Like
32

This: https://docs.cursor.com/context/@-symbols/@-docs

You can add Framework and APi Documentations. They will be indexed in cursor and then you can mention a specific doc for your task. In the Background its like a rag with domain specific informations about your favorite apis, frameworks and so onā€¦

33

Agreed 3.7 is bad for the reasons others have commented (likes to overengineer, tries to do too much, changes everything every time), BUTā€¦

Unifying Chat and Composer was AWESOME.

  • I can switch between modes in the same conversation. Previously, if I started chatting to plan a feature, I wouldnā€™t be able to switch to Agent to have it implement the plan for me.
  • Shortcuts (Cmd+I, Cmd+Shift+I, Cmd+L, Cmd+Shift+L) are now consistent. Previously, they were a mess.
1 Like
34

totally agree, The previous version was perfect, this one is not that good.

35

I stepped away for about 2 weeks and came back to a lot of changes - MCP, MPC servers, Sonnet 3.7, no composer. I appreciate all the work the Cursor team is putting in, but unfortunately itā€™s been pretty buggy for me too.

The ask vs agent works okay, although I canā€™t seem to change mid-chat. If I start in agent then Ask tends to still change things. 3.7 thinking seems more likely to ignore Ask v Agent and do whatever it wants.

3.7 thinking gives me wayyy too much response every time. I tried to plan something and it wrote and created 3 new documents on my first prompt. 3.7 regular seems to be better, but it still generates code snippets that I donā€™t want (to be fair, it was doing this with 3.5). And 3.5 is also failing to write to files or write in file formats regularly. In the summary of changes in Agent Iā€™ll still catch it writing code!

I suppose itā€™s possible that better rules would solve these problems. But after reading rave reviews about 46.0 (pre 3.7, right?) I was hoping for a better experience. Iā€™ll keep experimenting and Iā€™m sure devs will too.

1 Like
36

Feature creep at its finest.

I agree with the original post Composer goes off on its own and starts making changes that were not explicitly asked for.

Whats even worse is when it deletes blocks and blocks of working valid code because it had an opinion that IT Realized something was off when it wasnt.

How much of the ā€œAhh now I see the problemā€ or " Now I understand the issue" do we need to see?

2 Likes
37

Yep.

The latest couple of versions since introducing Claude 3.7 are absolutely horrible!

How do I go back the version before?

3 Likes
38

Nevermind, you can still download the previous .45 version here:

The latest .46 version is beyond unusable.
No matter how details I get with my commands, it does not follow instructions and completely destroys working files.

6 Likes
39

This new Cursor version just broke a lot of MCPā€™s Servers that I was using. In the .46.* some MCPā€™s arenā€™t getting their tools listed. Switching back to .45.* resolves the problem.

1 Like
40

I had no issues a few days ago. And the agent tool works great for debugging and testing etc! But yesterday and today it halts. Maybe the do with Cursor capacity to the LLMā€™s/models?
It hangs a lot and in php it creates a lot of simple not closing syntax issues { } etc etc due do halting the ai generating systems. So it corrupts my files now latest arm64 version mac

2 Likes
41

terrible as it auto updates 0.45 too so watch outā€¦