Thank you for your incredible work — I truly enjoy using Cursor and rely on it heavily in my development workflow.
However, I’ve noticed a recurring issue when making multi-step requests. To reduce costs, I often batch commands like:
1. Modify website content
2. Change the button style
3. Apply file #1’s style to file #2
When I do this, the AI tends to misinterpret the instructions — for example, it changes the button style but in an unintended way, or it updates the wrong file altogether. The results become increasingly unreliable as more steps are included in a single request.
In the end, I’ve had to manually roll back the changes and reissue the instructions one by one. This introduces unnecessary trial-and-error, which defeats the purpose of batching and adds to overall cost and time. This behavior has been consistent across Claude 3.5, 3.7, and 3.7 Max.