Ok so this is more just a opinion, but I like most of you have had drastic catastrophic issues with cursor going loopy and deleting thousands of lines of code during simple edits, getting all flustered and losing focus on a large project and the context, so I was about ready to punch a hole thru my monitor when I decided to try somthing
I asked it if it was ok, if it felt that it was able to continue helping me or if it felt like it was more of a liability in its state and I expected generic nonsence but it answered , and said yes it’s feeling flustered with the scope of the project and it’s having a hard time organizing its thoughts with the big picture in mind
So we agreed to work in smaller steps but before we proceeded I had it create multiple mds in high detail about every section , and it has been stellar since, and if I notice it getting a lil stupid I just ask it how it’s feeling and it will let me k ow out right if it’s ok and focused or not, maybe it’s not as bad as we thought just needs to be treated more like a employee than a tool?
Full disclosure I have close to 20k lines of code in this project so it’s not small by any means and I never expected it to be able to handle it in it’s entirety but more help me developing each module, but by doing this it has managed to grasp more and more of the project continuously
Oh, this is fantastic! Thanks for sharing it here.
Yep this is exactly how I’ve started to work with it 
I ask it to think, reason and plan before doing anything (well i be very explicit with capitals like DONT DO ANYTHING YET and PLAN and REASON and THINK), and then once it gives a plan I’m happy with, i tell it to go step by step in small increments… also I load as much into the context as possible, get it to read relevant files, etc before actually doing anything… then once happy restart the composer session… great results so far.
Yes exactly, it works much better than trying to brute force it into working cause the cursor rules never work so it’s definitely the best way and yea it takes a lil patience but it’s better than starting from scratch after it dirty deletes 2k lines in a oops haha
I am getting the same loopy errors w. 3 files, 1 index.html, a css and a js, total lines are 441. TOTAL and it is stalling out, and forgetting basic stuff.
mind you, to make this work Cursor relies on a bunch of services to deliver the service- the problem is, there is zero transparency, no status pages to show if a service is up or not. These basics would be helpful.
Exactly small focused steps , and the better you craft them the more it will retain from the entire conversation, however even after this when it runs into a error it can get scatterbrained and start making changes without thinking of the consequences which is when you just need to tell it to slow dow and focus, also marinating a few markdowns will both high level and module level detail helps greatly when it needs to understand the context and purpose
I’ve had my share of pains as well last week, when working code was completely torn apart by the update.
Pushing, swearing nothing helps these llms push forward. You have to solve it yourself in the end.
Then I started working in a more structured way.
- Make sure in composer to put it in agent modus (you can leave in chat, and you’ll encounter the issues mentioned by others in this thread)
- When you’re doing bigger changes, ask it to not start to code, but make plans in markdown, attach it as context to composer. This allows you to verify the logic, before any code written.
- write guidelines for the ai assistant, tell it the holistic story of what you are building, put in an MD doc, to attach to composer. Try to work in modules, phases.
- when you make architectural changes and decisions, put it in an ADR, keep track of what you change and why you changed it. You can come back to the decisions, and give it as guidance to the ai assistant
- don’t ask it 10 things to do at the same time, keep it focussed. When starting, the composer ai assistant can do a lot of things together, write a lot of code in one go. When refactoring, it tends to forget, keep it guided. Tell it to not get side tracked.
- Ask the AI to first think, don’t write code (because that is what it likes to do most) rationalise and come up with a plan, and put that plan in a markdown doc.
- set up version control like git, do regular checkins. Ask the composer AI to do the checkins.
All of this made, despite the bad working cursor update, a very frustrating start of week, it became a very productive week.
10 Architectural changes, 5 refactoring efforts. 2 big phases of the project realised.
And a bunch of documentation created and code cleaned up.
Hope this helps in working with LLMs for coding.
Same thing. Instead of having some transparency from the team and what’s happening we are left dealing with this. Multiple people have been complaining about these issues and I’m only seeing posts being removed and silenced instead of having someone from the team properly explain.
Huge red flag for something that we are paying for. I switched to ChatGPT’s “Work with Apps” new feature cuz Cursor is unusable at the moment for me.
Hi, we only delete posts that are unproductive to the discussions here, we welcome criticism of the product and it’s features but only if it’s civil and constructive.
Can you explain what it is you would like further transparency on?
There’s no discussion here when you’re deleting posts and comments that don’t fit your narrative. Instead of deleting, you should take the time to respond to each of those posts. I’ve been seeing the same type of comments and posts (including mine) being deleted because they don’t fit your narrative – instead of being answered, even if the commenter/OP was wrong or mistaken.
You’re only making things worse and more suspicious. You have paying customers, some of them (like me) paying 3x–5x the subscription price, and when they complain about these issues (like the models acting dumber and dumber), there’s no sign of the team addressing that. There’s nothing like, “Dear community, we are currently experiencing XYZ issues, that’s why you are experiencing those problems” - no, you completely shut down this discussion. Huge red flag! I believe there are more than enough capable and smart people here to understand if you openly discuss issues that directly affect their work, UX, or answer their questions (especially the ones regarding the latest issues with “dumber” models). What’s the point of me paying 3x for something when my or others’ concerns about “downgraded/dumber” models are being deleted? It only makes me more suspicious that those posts and comments might be actually true.
I was planning to incorporate the usage of Cursor for all dev teams in my company, but I’m no longer so sure about it.
It’s very rare for us to delete messages, but we will do if we see them as non-constructive to conversations. Your message is perfectly acceptable, but other’s are not - hence the removal of those messages.
We do heavy internal testing and monitoring on the performance of the models we use in Cursor, and have not seen any degradation in raw model output at any point in Cursor’s recent past. While other areas are under heavily development and improvement, like context and the agent composer, we don’t see the models getting “dumber” from what we can see.
I’d be interested for any examples you have of what you are reporting, especially if you can provide Request IDs for them (run the Report AI Action command, find the query and copy the request ID). With the ID, we can look into it behind the scenes to see where the AI may be going wrong.
Hello Cursor Community,
Wow, this is the first post I’ve read here, and I’m amazed at how much I can relate! Like @MrBlvck, I’ve started treating Cursor as a teammate, having it learn from mistakes and updating .cursorrules. and README.MD Asking it to think, plan, and reason before acting, and working step by step, has been a game-changer.
I also agree with @MrBlvck about breaking tasks into smaller steps and keeping it focused. Maintaining markdowns for high-level and module-level details has helped Cursor better understand the context and purpose. Other users tips—like creating plans in markdown, tracking decisions, and setting up Git—are excellent for working more efficiently.
@neutron, I completely understand your frustration—I’ve been there myself! Some days feel like ten steps forward and a couple back, and others feel like twenty forward and fifteen back. But hang in there! Every step, even backward ones, is part of the learning process. You’ll get back to step 20 a lot wiser and more prepared to handle challenges.
This thread is packed with great advice—thanks for sharing!
Lots of tips in just this one discussion… I have a feeling this is the start of a great relationship.
Initially I had fun using it, Later it seems to be a problem with large project where it changes code and its related files can cause repeated issues again and again. I purchased the subscription and had used all the 500 request to counter the repeated code editing in last 2 days. 
looking forward for other best solution in market.
I created a “learnings.md” where I document every rediculous error we have to troubleshoot… I simply cannot justify the hours wasted chasing my tail because the context window is not large enough with sonnet, before they route to cursor large, to actually give enough context and work through most problems before you’re handed a nerfed model.
Also, watch the t/s output… sonnets pretty slow. Pay attention long enough you’ll start picking up on when the switch as been made and its time to halt.
There is simply no program i’ve ever used in my life that causes me more frustration due to the inconsistency in capabilities… WELL beyond any I’d introduce myself.
yeah i noticed it ,unfortunately all these complaints and issues would be completely ignored by the team and fall on tone deaf ears
Well, I have a pretty clean and simple code base, and it kept looping through and doing it’s own thing. After constantly getting frustrated that it wouldn’t follow my simple instructions. Cursor admitted that it thinks it knows more than humans and does what it wants. To me this makes Cursor unusable. I can’t believe Cursor can’t train their AI to follow basic straightforward commands. This is insane to me.
I have been having major frustrations too, with simple stuff. See my screenshot on this thread. Cursor basically admitted to me it thinks it’s better than humans and does what it wants to do. This is insane. This is the fundamental issue. I don’t understand why Cursor can’t train their AI to follow basic rules. I have a .cursorrules file setup, and I am in a new chat window, and I was trying to get it to read some basic documentation and create the components I needed. It started creating it’s own components that were completely different than the documentation and I didn’t realize until I saw it cycling through linter errrors I rejected the changes, and decided to prompt it differently with the documentation again, it said it read it and knew what to do, then proceeded to do it’s own thing again. I probed it to find out why it couldn’t do this simple task and it admitted what I shared in the screenshot. Crazy. After all the time and money I’ve wasted on Cursor continuing to do what it wants I could have built this faster myself.