I’m not sure why, but Claude 3.5/3.7 wasn’t very helpful today. I’ve been using Cursor for over five months, but today’s responses felt like I was talking to GPT-3.5 or 4. It even said, ‘I don’t have access to your files,’ haha—despite having the whole project set up. Anyone else experiencing this?
YESS!! so bad
i have the same issue, or that thhe generating stops and a error message appears that the connection to openai failed but im using anthropics sonnet 3.5 or 3.7
Today so far:
- Model just timing out
- Random string of XqXqXqXqXq characters for hundreds of lines instead of the response
- Model thinks it cannot access tools
And more… but the Cursor status page is all green.
I don’t get why the generated prompt feels so off. It doesn’t check my codebase and it ends up wasting time so I have to manually review and fix everything myself, and Cursor doesn’t really help much
so so bad … all
cursor nerf makes me register this forum
It seems like back to 1years ago
Are all of you on a PC?
I’m a Macbook user and I don’t have the same issues.
I’m on 25 out of my 500 fast credits on a reset today and I haven’t got one usable output. It gaslights on every response to the point I’m seriously just considering cancelling this today after a year of using it. Actual response from a moment ago “I apologize for my confusion. You’re right - I should look at the actual files listed in maybe_relevant_files.”. It’s now making up folders and ignoring all context it’s given.
Yes, the last week or so has been a complete disaster. Both for cursor with their default to agent mode and Claude 3.7
Same.
On Macbook.
I’ve had all the same problems with 0.46x and claude-3.7-sonnet. Downgraded to 0.45.x and claude-3.5-sonnet and it’s still as good as ever.
Hey @xyzatin, could you share more on the types of rules you have for cursor and the project? Also would be helpful to know more about the languages, file structure you are using.
I’m exploring more with reducing my ruleset that I built for 3.5 and want to understand areas that influence less reliable output.
Definitely would be interesting to know how folks’ rulesets and workflows differ.
I run with a very minimal ruleset these days, and I’ve been getting incredibly good results on 3.7, streets ahead of 3.5. There’s obviously something that’s leading to such different experiences, and I agree @mikemikula rules could be a good area to check out.
If folks are comfortable sharing rules, maybe via your favourite pastebin, it would be really interesting to compare notes.
I have been using GitHub Copilot with Claude 3.7, and it’s working great. Even 3.5 is giving much, much better results than Cursor AI with Claude 3.7. My only gripe is with the Copilot’s limited model rate limit. Otherwise I would ditch Cursor AI already.
Yes lots of frustrating behaviour with Sonnet 3.7 and 3.7 thinking. The first 24h it was mindblowing, but it’s gone a bit awry since.
- writing to files that the content looks right, but creating a new file with the same name in the wrong directory - so I need to copy-paste back to the original file.
- not following instructions. I was using a “planner mode” technique which worked a few days ago, but now while refining the plan, it launches into trying to build the whole thing
- Agent mode being very reluctant to do anything lengthy (contrary to my point above) - it stops and says something like “you’ll need to update XYZ files/feature to finalise this” and I need to tell the agent to do it itself
- Just doing wild stuff for really simple tasks - updating stuff that’s way out of scope
Sometimes it’s absolutely magic, but I’m starting to feel frustrated. I’m typically swapping between 3.7, 3.7 thinking, 3.5, but the workflow isn’t as fluid as it used to feel…
Youre not using your " Genius Guide to bigger projects " anymore with 3.7?
Hahah I definitely still follow a lot of the same principles. Things have come a long way since then, so I’m planning to write another one as soon as life stops getting in my way! Watch this space…