I’m new to cursor, so this has probably been covered before, but searches haven’t turned up anything useful.
Do we have a way to prevent the agent from making changes that aren’t approved?
I love the power of this tool, but it quite often starts going berserk, making changes all over the place, in areas completely unrelated to the current request, breaking things as it goes.
Google says there’s a YOLO mode in the settings that can be disabled, but I don’t see that option anywhere.
I’ve set rules that explicitly state that it must submit it’s suggestions for review, and not make any changes without approval, yet it seems that I still have to remind it every 2-3 messages.
It’s getting tedious having to write Do not make any changes, just provide your suggestions" with every request.
I’ve been having the same issue. I might write a routine on the side and get it the way I want, then add it to the project and tell cursor to use it for X. Then Cursor takes over and solves the problem a different way. So I asked Cursor how I could protect the code I introduce, and it added the following to my code (will see if it works):
/*
PROTECTED FILE - Do not modify without explicit permission
This PageTable implementation correctly parses tables in images
@okkine, might be a late response, but if you’re still struggling, here’s my comprehensive solution.
Quick Settings Access
First, access your Chat settings through:
Settings Icon → Chat, or
Ctrl+Shift+J (Windows) or Cmd+Shift+J (Mac) to open settings quickly
About YOLO/Auto-Run Mode
You mentioned Google says there’s a YOLO mode - Google is giving you outdated results there. They’ve renamed it to “Auto-Run Mode” and you can find it in Settings → Chat → Auto-Run → Auto-Run Mode.
Here’s the thing though - Auto-Run mode only controls whether the agent asks for permission before running terminal commands like npm install or git commit. It doesn’t actually control code editing permissions.
If I understood you correctly, your issue is about code editing control, not command execution, so this won’t help with your problem of unwanted code changes.
Here are my suggestions:
1. Use Ask Mode Instead of Agent Mode
Ask Mode: Read-only, perfect for getting suggestions without any changes
Agent Mode: Full autonomy (what’s causing your problem)
Switch to Ask mode when you want suggestions without automatic changes
2. Custom Modes - The Best Solution
The best approach is to use Custom Modes (BETA):
Enable Custom Modes: Go to Settings → Chat → Custom Modes (BETA)
Create your own custom agents: You can basically build your own versions of the agent that only do what you want them to do
Fine-tune Permissions: You can control exactly what tools each mode can access (Search, Edit, Terminal, etc.)
Example Custom Mode Setup:
Name: “Review-Only Assistant”
Tools: Enable Search, Disable Edit and Terminal
Instructions: “You must always ask for explicit approval before making any code changes. Provide detailed explanations and suggestions, but never modify files without permission.”
3. Manual Mode: Only edits explicitly selected files
You need to open the file you want to work on, then use chat
It can only modify the file you’re currently looking at
It can’t access any other files unless you use @ and attach the file to it manually
Additional Control Methods
Cursor Rules (.cursorrules file)
Add rules that modify the AI’s behavior without needing to prompt it repeatedly. Create a .cursorrules file in your project root or go to Settings → Rules & Memories and add rules like:
- Always ask for approval before making any file changes
- Provide explanations for all suggestions
- Never modify files without explicit permission
Make sure to set this to apply for always.
Rules in Settings
Go to Settings → Rules & Memories → User Rules and add rules that apply to all interactions.
Memory Feature
You can tell the agent to save in memory specific preferences like “save in memory that I always want to review changes before you apply them.”
