Is it just me or does cursor and other programs like it. undo all the work you have done just because it makes a complete ■■■■■ up of the script.
Without programming knowledge i cant fix myself. but you half expect it to get it right as its also a paid service.
3 Likes
Right!?! I feel your pain! This happened to me 3x - and the agent went into my backup folders and replaced both backups with a template placeholder code, essentially wiping out all of my work and backups - had to start from scratch again.
Now I can tell when to stop it before it starts its destructive behavior and pattern of going in circles and creating more issues with random solutions it tries to apply. They key is when it can’t fix the problem, tell it to try a different approach, library, or method. I always ask to verify if the solution it is trying will fix the problem.
As frustrating as it has been, it has been a good learning experience that i has been beneficial and when the Agent is working properly - we have done some impressive work. So far the benefits still outweigh the negatives…
Good luck.
1 Like
Thanks for your reply. yes the benefits do outweigh, it seems to have problems when the codebase gets too large and it seems to keep creating more files and loses track of what its doing. then makes more mistakes. reverts back to backup or makes large changes and ends up breaking the code completely. I even go the bot to edit a file to keep itself updated on progress and make notes on how it fixed certain parts. But even with this, its still challenging for me and the bot. Do you have any suggestions that you have managed to get working successfully?
Many thanks
These tools are awesome but nowhere near perfect.
Are you using source control with a github backup?
If I don’t like the path claude 3.7 is taking, I’ll revert back and try again.
Sometimes when it’s in a fix - error loop, I’ll keep going and it eventually figures it out. Otherwise, one click and I’m back to the last known good code.
Could you please explain source control, and where to enable it. Also, what will it do to benefit? As right now, my current pathway doesn’t seem to fix it, just makes it worse. the older versions would be ok on this. but there were other issues. every version has its own set of problems. Right now i’m using the most recent version. I don’t use github anymore as it added my api keys to github by hard coding the keys. Another issue i just found, the bot started installing dependencies from the project i did prior. Which have nothing to do with the current project and are in completely different folders.
I’ve been slowly adding cursor rules to prevent this sort of thing. Telling it to review the existing solution and iterate on it rather than adding new files. Telling it not to change the ports we’re using in the launch configuration.
But yeah, I feel your pain. It’s like painting with a fully-automatic shotgun.
1 Like
I tried adding rules, but i couldn’t get it to work. so i created a file for it to follow, then after a certain time, it would forget.
i tried, but not sure how to add it. turns red. am i getting this all wrong. i tried different working. i don’t get it
I also just use Git source control. Once you get your project to a good base line I’d suggest creating a private repo on GitHub and then push your project code to that repo. After that start making incremental changes to your project with Cursor that way if something goes wrong you can simply discard the changes and start a new chat or you can even revert to previous revisions in your source control.
The rules also are a good suggestion but I have not used rules that much as yet. I do often create a file with all the rules I want it to follow (coming from copilot old habits die hard) and you can simply add that file into the context of your conversation with Cursor telling it to reference the rules or guidelines in a file you created such as guidelines.md. That way it will follow the instructions in that file or at least will try to, it is AI after all and sometimes still will try do things you don’t expect.
3 Likes
Start with just initializing a repo in vscode in the source control tab. This makes it easy to revert code. I use github mostly for backup but occasionally I’ll start over, delete the project and pull my repo if things get too messed up. Also, I use rules to let Claude know basic stuff like the language and framework plus where to look for files that aren’t in the project like if you have separate client - server code.
Remember, this AI coding stuff is pretty new. The tools will improve.
1 Like
It doesnt always follow rules though!
Don’t trust 3.7 or thinking modal at all, still perfectly works 3.5.
3 Likes
Honestly, I’m about to ditch Cursor for all kinds of reasons.
Yes, the AI will go rogue and break everything or overwrite things that are working fine for no reason.
It has no contextual memory past the last message.
I have been writing super thorough prompts and have a huge CONSTRAINTS section. It’s been better but inconsistent.
At first it feels like superpowers but I have just spent three days trying to fix errors Cursor and Sonnet 3.7 introduced. So, I had three days of acceleration and now three days to backtrack, refactor, and have manageable code from the thing we did together.
The only thing I am convinced is 100% better is that I don’t have to spend half a day looking for a random typo.
2 Likes
Yeah, strangely 3.7 thinking is the biggest airhead and is doing the most stupid egregious things. Like working with DOGE.
1 Like
It’s worth noting that I saw on YouTube there are videos covering a new feature called “Cursor rules” and I haven’t gone through them yet but I’m pretty sure that’s where you could potentially limit the damage and focus on accelerating.
I should get over my frustration and say that it’s a huge time saver on boiler plate code that can boost mad productivity if you manage it carefully and keep your own awareness of the overall architecture.
I also have been really purposefully forcing 3.5 sonnet to use DRY and Single Source of Truth principles, forcing it to use modular code through imports and exports (it’s painful and slow) and it’s been much better today at not creating files with 600+ lines of code that turn out to be totally redundant.
Maybe we should just be more deliberate and careful and grateful. Still get so frustrated I feel like blowing it up though.
If you have absolutely no coding skills or programming knowledge, I firmly believe that learning and utilizing programming will lead to faster results than relying solely on Cursor.
1 Like
This is where the mpc memory server really shines.
+++
claude 3.7 feels much dumber
It changes extra code even when I ask to stay from other changes
I totally agree. My programming skills have only improved since using the AI coders. Just asking the genie to grant the wish of a full blown app in one prompt isn’t there yet.
1 Like
Sadly, but yeah, that’s sometimes true.
The cost of 2 requests doesn’t always match the output, it can still be a bit off, hahaha.
I’d rather go with DeepSeekv3, it’s fast, powerful, smart enough, and free!
1 Like