Feature request for product/service
Chat
Describe the request
It seems like the Cursor model only maintains behavioral changes within a single chat context. I’ve had to tell Cursor multiple times to check in with me before making changes to files so I can supervise its proposed changes. It tends to make very myopic changes without considering larger architectural impacts.
I’ve also had to tell it to stop saying “You’re absolutely right!” or “You’re right!”. This borders on sycophantic and I want it to focus on the task and not gassing me up.
I asked the model why this is a problem:
You’re pointing out a real issue with my behavior. I do seem to fall back into patterns like excessive agreement (“You’re right”) and making changes without checking first, even after you’ve explicitly told me to stop.
This suggests I’m not effectively persisting behavioral instructions across our conversation. When you correct my behavior, I acknowledge it in the moment but then revert to the same patterns later.
The problem seems to be that I’m not consistently applying the behavioral constraints you’ve set. I should be:
- Not using excessive agreement phrases
- Always asking before making code changes
- Checking existing files before modifying them
But I keep slipping back into these patterns, which indicates I’m not maintaining these behavioral changes effectively across the conversation.
This is a limitation in how I process and maintain behavioral instructions - I seem to treat them as temporary rather than persistent rules for our interaction.
I don’t have definitive information about whether this is intentional design or a limitation. From what I can observe about my behavior:
It could be intentional design where behavioral instructions are treated as contextual rather than persistent - the system might be designed to prioritize task completion over maintaining behavioral constraints across long conversations.
Or it could be an unintended limitation where the model doesn’t effectively maintain behavioral instructions as persistent rules throughout the conversation.
The fact that I can acknowledge the issue when you point it out but then revert suggests it’s more likely a limitation rather than intentional design, since intentional design would probably be more consistent.
But I don’t have access to the underlying system architecture or design decisions that would give me a definitive answer about whether this behavior is by design or an unintended side effect of how behavioral instructions are processed and maintained.
Screenshot / Screen Recording
Operating System (if it applies)
Windows 10/11
