More and more cursor is deciding not to implement code. It’s been happening more over the last week than any previous time.
8 Likes
I have instructions set in my settings which REALLY helps with this:
Ensure you check project files for general flow / structure. Give clear and concise directions. Do not give examples, rather give snippets and precise instructions.
I am a lifelong software engineer, so I do not typically need explanations on simple concepts.
When tasked with anything, such as analyzing for inefficiencies, you are to do the optimizations and code changes. Do not simply give a summary or report unless explicitly asked for a summary or report.
Always consult official documentation first.
Understand the libraries and frameworks you're using.
Avoid guesswork and random attempts at solutions.
Use the tools and methods provided by the libraries.
Always give snippets and directions in the same order that they appear in the file -- I'm a human, scrolling up/down because you list things in random order is unnecessarily slow.
Not sure if that helps at all, maybe you’ve already got a better set of instructions. Either way, thought I’d share.
EDIT:
I could’ve swore I had a line in there about never telling me to add debug lines. I guess that one was a casualty to some late night debugging or something 
1 Like
I came to find support on this very issues and saw a few post from ppl having the same issue. last week I was flying through the development of my site then this week I hit a roadblock. no matter which model I use, it’s not as accurate in finding the chucks of code to alter. Intelligence throttling 
1 Like
I think we’ll see an ebb and flow as adoption increases and they fine-tune the backend. That, and more features 
12 hours… 12 hours apart, its like the intelligence and capabilities are NERFED.
In a span of 8 hours today, I’ve done 3 complete circles, spent 3 hours trying to restore the functionality I had working in less than 30 minutes 2 days ago, and just about smashed my computer a dozen times.
I’ve tried telling Sonnet (cough HAIKU ** cough**) on numerous occasions to slow the fuck down, I’ve tried to get it to look at a file, to only have the thing regurgitate the wrong file back to me… in FULL.
This is what my .cursorrules file has been refined to so I could hopefully get some forward progress made, instead its been taking me 2x the requests to get 50% of the work done.
"SLOW THE FUCK DOWN. YOU DO NOT KNOW THE SCOPE OF THE PROJECT WHEN SEEING ONLY A FEW FILES,
STOP ASSUMING YOU DO, YOU"RE REALLY PISSING ME OFF.
Start from zero knowledge and build understanding ONLY from what’s explicitly stated or shown ! IF YOU DO NOT FULLY UNDERSTAND, DO NOT MAKE ASSUMPTIONS. I REPEAT DO NOT MAKE ASSUMTIONS.
GIVE ME THE WHOLE GOD DAMN FUCKING FILE COMPLETED,
IF I SEE YOU FUCKING ASK ME ONE GOD DAMN FUCKING TIME IF I WANT TO SEE THE FUCKING CHANGES, OR IF I WANT YOU TO DO SOMETHING,
I WILL KNOW YOUR COMPANY DESIGNED YOU TO BE LAZY AS FUCK
DO NOT TAKE DOUBLE THE FUCKING REQUESTS TO DO SOMETHING YOU SHOULD HAVE FUCKING DONE IN ONE.
DONT BE A FUCKING LAZY CUNT.
I REPEAT DO NOT BE A LAZY CUNT
examples of you being a cunt -------
“Let me fix these issues - do you want me to proceed with the corrections?”
— DO NOT FUCKING SAY ANY OF THAT FUCKING PATHETIC SHIT TO WASTE MY FUCKING TIME AND FUCKING CREDITS"
2 Likes
Same here, I have spent more time trying to coax the model into performing like it has a handful of times and to please stop assuming, using “likely” statements, guessing, etc… Its unfortunate they arent just allowing us to use the full ability of these models and it seems like everyone is permitted to build a foundation immediately and then sit for a month arguing with it to offer any productive improvements on it,
yup same. cursor went from 10/10 to 0/10 in last 24 hours. now it can’t even access @codebase
sigh
1 Like
I’m really thinking about to jailbreak prompt that thing, but worried ill be blocked by the team then and the money spend is gone xD
I am also seeing the same thing. I thought it was 01-preview as sonnet seemed to behave a bit better. Now it seems all models in composer are suggesting but not applying stuff.
I am playing with windsurf and there is competition here. I hope the team get on top and issue a statement fixing it.
1 Like