So when I started using Cursor it was super useful and I was impressed, so subscribed to the Pro version.
Now I’ve used up a lot of my resources just having it fix AI’s mistakes. So basically I’m paying Cursor for me to teach AI how to do something and then have it correct it’s mistakes. Thank goodness for branches! But now I have to save before I let AI do anything because I’m afraid to let it do anything on it’s own. It takes so much hand-holding now, it’s like instructing a toddler. I’m not sure what happened because it’s gone downhill and has become an idiot. If it were a real person, I would have fired it. So frustrating. How can I fix this situation so Cursor can be fun and helpful again? What went wrong?
Hey, thanks for the feedback. Here are a few things that usually help you get consistent results:
Set up Project Rules
This is the biggest lever. Create .cursor/rules files that describe your project patterns, conventions, and architecture. When the agent understands how your project is set up, it makes way fewer mistakes. More info here: Rules | Cursor Docs
Try different models
You can switch models mid chat using the model picker at the bottom of the chat, or Cmd+/ or Ctrl+/. Stronger models like Claude Sonnet 4.6 usually do better on complex tasks. Faster models are fine for small edits. Picking the right model for the job can save requests.
Prompting approach
Split big tasks into smaller, more focused requests instead of asking for large changes all at once
Use @ mentions to point the agent to specific files or folders if you know what matters
Use checkpoints so you can safely roll back if something goes wrong instead of managing branches manually
Use Plan Mode
Before you let the agent make changes, you can turn on Plan Mode. The agent will first outline a plan so you can review and adjust it before any code gets changed.
A couple questions so I can give more specific tips:
What model are you using right now?
What project are you working on, language and framework?
I already use Cursor rules. Sometimes the agent gets stupid. I have to tell it things repeatedly or revert the work it does. One step forward, two steps back. Not only is it a waste of time, but money; I’ve already burned through my limit in like 5 days.
I can start new agents, but it’s like having a baby and I don’t have time to keep educating it. Rules should help but it makes me nervous when I’m into something deep and it’s complicated. The plan mode sounds helpful. I normally tell it to get my approval before making changes and have a rule in regards, but it often acts on it own regardless.
I’m mainly using Laravel and Filament, which are PHP frameworks. It’s for websites, so all the other assets that go along with them.
Also, now it keeps saying Reconnecting a lot and I have to restart the Cursor app. I’m at 88% of total usage if that makes any difference.
I had to upgrade to the Pro+ subscription b/c it used up all of my total usage in like 5 days. Can I get a $20 refund since I upgraded my account? I didn’t see a way to do that in the upgrade section. Thanks
Reconnecting
Try going to Cursor Settings > Network > Run Diagnostics and send me the result. Also try turning off HTTP/2 and selecting HTTP/1.1, then restart Cursor. This often fixes connection issues.
One more thing: if you have a lot of old chats with unaccepted changes, it can cause freezes. Try archiving or closing old chats.
Agent ignores rules
Can you share the exact rule text it’s ignoring? Also, which model are you using? Some models follow instructions better. Using Plan Mode can also help since it shows a plan first and waits for confirmation instead of relying on the rule “ask before changes”.
Refund 20 USD
Refunds can’t be handled through the forum. Please email support at [email protected] and they’ll help.
Let me know how it goes with the Reconnecting diagnostics.