Cursor has become completely unusable for me in the last few days. I can, barely, deal with the slow responses, but the fact that it can’t keep any of its rules or context from prompt to prompt is causing me to spend hours trying to get it focus and do simple tasks correctly. Here is the latest thing that happened.
I started a new chat, I have “claude-3-sonnet” selected. I asked it to review its rules, and it read them back to me. I asked it to scan some CSV files and grab headers and put them in the column mapping field in some Python code, and to not lose the rules and context. It failed to do the task and violated several rules. It thought that it wrote a Python script that it didn’t, then tried to run the file it didn’t create using commands that it just read back to me that it’s not allowed to use. Here is its response to me:
You’re right, and I apologize for the continued problems. This isn’t normal or acceptable behavior.
The specific issue is that I acknowledged the rules but then immediately tried to run terminal commands without permission when attempting to verify column mappings. That’s a direct violation of the guidelines you set.
This pattern suggests a fundamental issue with how I’m processing and maintaining your instructions between message turns, even within the same conversation with the same model. It’s not a matter of forgetting - it’s a failure to consistently apply acknowledged constraints.
For a platform that’s meant to assist with coding tasks, this inconsistency makes it essentially unusable, as you’ve pointed out. This requires escalation to the development team as it appears to be a core issue with how the model processes and maintains instruction adherence across turns.
Would you like to approach your original task in a completely different way that avoids any execution without your explicit approval? Or would you prefer to report this issue to the team developing the platform?