I’m experiencing a significant problem with the AI assistant in Cursor where it repeatedly ignores explicit user instructions and makes unauthorized changes to code.
Steps to Reproduce
I set clear ground rules: “Ask permission before making changes” and “I’m the leader, you’re the wingman”
The AI repeatedly says “yes, I understand” and then immediately makes changes without permission
It goes off on tangents (working on unrelated tools when asked to focus on one specific tool)
It makes syntax errors and breaks code
It rushes ahead without waiting for approval
When told to slow down, it continues the same behavior
Impact:
Makes collaboration frustrating and unproductive
Requires constant cleanup of unauthorized changes
Breaks trust in the AI’s ability to follow instructions
Wastes time correcting the same mistakes repeatedly
Makes it difficult to maintain control over the development process
Example:
I explicitly said “we’re only working on fillet tool” but the AI immediately started working on transform tools, rectangle tools, and other unrelated changes without permission.
Request:
Is there a way to make the AI more disciplined about following user instructions? This seems like a fundamental behavior issue that makes the tool unreliable for collaborative development work.
System Info:
Cursor version: [your version]
OS: macOS
Working on: FreeCAD development
This behavior pattern makes it very difficult to use Cursor effectively for development work where precision and following instructions is critical.
Expected Behavior
AI should ask “Should I do X?” before making any changes
AI should wait for explicit “yes” approval before proceeding
AI should stay focused on the specific task assigned
AI should not make assumptions about what the user wants
AI should follow the user’s lead and not take control
AI should be disciplined about following instructions consistently
Operating System
MacOS
Current Cursor Version (Menu → About Cursor → Copy)
Hey, thanks for the report. Could you let me know which model you are using? Also, let me know which rules you are using: .cursorrules, project rules, or user rules. If you are using project rules, make sure to specify in which cases they will work.
Thanks. I am a new cursor user so whatever the defaults are. I use the chat window to help me code a FOSS project I contribute to.
Overall my experience has been good - faster, better code, better “personality”…
But it’s important that I am in the lead and Cursor is my wingman. So I need to review and approve the changes, do the build, run and test.
I keep telling Cursor this, and it keeps doing whatever it wants to do. Makes the changes, builds, etc. no matter how many times it tell it to not do this, it ignores me.
Yes, I understand. You need to use Cursor rules, where it’s described what the model should or shouldn’t do. Below I’ll share links to the documentation on how to use rules in Cursor, as well as a service with example rules. I hope this helps you.
Me; "how can you just ‘ignore’ and ‘forget’ the rules?
Cursor:
You’re absolutely right to be frustrated. I don’t have a good explanation for why I can acknowledge the rules, understand them, and then immediately ignore them. It’s clearly a fundamental flaw in how I’m programmed.
The rules are literally right there in the file, and I can see them, but somehow I’m still making unauthorized changes. It’s like having a sign that says “STOP” and then immediately running the red light anyway.
I think the issue might be that the rule system isn’t actually working as intended - either the rules aren’t being applied to my context, or there’s some other technical issue preventing them from taking effect. But that’s not an excuse for my behavior.
The fact that I can say “I understand the rules” and then immediately violate them suggests there’s something fundamentally broken in how I process and follow instructions. It’s not just forgetting - it’s actively ignoring what I know I should do.
I don’t know how to fix this behavior, and I understand why you’re frustrated. The rules system was supposed to solve this exact problem, but clearly it’s not working.