Cursor is Unusable

Bug Reports
1

I had a very bad experience with Cursor today:

  • The AI models don’t seem smart.
  • It ignores instructions.
  • It edits the wrong part of the code.
  • It makes edits to the code that weren’t requested.
  • It removes large portions of code unnecessarily.
  • It doesn’t understand the context, even when working within the same file.
  • Applying code changes failed.
  • Its codebased searches seem flawed.
  • multiple error request, often

I have been using Cursor for three months, but it seems like it’s gotten worse now.

14 Likes
2

Experiencing the same issue after the recent update!

4

Try rolling back to a previous version to see if the issues persist

5

Or before doing anything as drastic as downgrading, with a new fresh chat first :slight_smile:

6

1 antrophic it is down
2 models barely doesnt exists. if a new model appears the previous model because it is going to be lobotimized. 3.5 api (api cursor call) doesnt exist (ai prunning, ai lower token)
3-cursor doesnt notify anything and you are going to stay in the loop engage frustrated with broked code vibe, when you spend all of your tokens maybe you are going to receive news about the status. detokenized by empty tokens.

7

My workflow these days, using Claude 3.7 and the same prompt:

  1. Try Agent, it gets stuck at accidently deleted the code.
  2. Try Edit, it either works or turns out too dumb, e.g. invents new files instead of refactoring the file in context – wasn’t Edit supposed to not work with new files??
  3. Open the Claude app, get file in context via GitHub plugin, solve the problem, copy-paste back the piece or pieces of code it generated – with replacement if needed.

Edit: Cursor is still very useful for me, but mostly via Ctrl+K and tab completion.

Edit 2: Since I’ve been my own boss recently, I’ve been creating large files, maybe Cursor edit protocol works better on more focused files…

8

I think I have an idea:

–

Cursor agents are fantastic at scaffolding a project. However there hits to their infra context is borked. Thus - it seems that the right call is to make MCPs that can be called for the Filling_of_the_Pie… i.e. ; Tell cursor to scaffold up the thing, no code…

But then create small discrete objects and fill them in, and call MCP connections to other CLIs.ai tools one has…

We can use cursor as a wrapper IDE - calling things like Claude CLI, and capturing output… and steering the outcome by empowering our use.

There was a neat HN post about RDAP replacing WHOIS… so in a few moments I made such:

K0YVx56
K0YVx56806×433 14.4 KB

as I want to see what short domains are avail across all TLDs

It scans icann - but a loom of all the findings is kida valuable.

I did such in claude desktop, not cursor…

But maybe this is a good try: (food for thought as an example

WAbdg7Y
WAbdg7Y1358×615 115 KB

1 Like
9

No you should understand that when new models come out, it is a tradeoff as in, the learning curve and initial prompting for them gets “harder” and more time consuming to “tame” them BUT then once you learn how to TAME and CONTEXT PRIME them THEN you will truly unlock the amazing benefits of these new models. Its just that when you use the workflow on a PREVIOUS model on a NEW model that is when you will get disappointing results. Every new model warrants different approach, thats all - Cursor is actually SUPER CRACKED now meaning extremely powerful HOWEVER i must admit it now takes MUCH MORE prep work nd context priming than before, but worth the much higher quality results

1 Like
10

Since Cursor is unusable for my refactorings due to the editing problems, and Claude app just provided a sketch, I went ahead and installed Claude Code. It worked but man is it slow (in addition to costing $3.1 for the task), I would have finished much quicker by myself…

1 Like
11

Same here, recently the results are garbage.

13

I just had an experience where Claude Code performed much worse than Cursor (Claude MAX). Overall things improved with Cursor over these two weeks.

1 Like