I am a subscriber of Cursor, which is a powerful tool that currently no other IDE can surpass. However, I believe Cursor is deteriorating in quality, and I think it is time to voice my concerns:
Issue One: Problems with File Handling
The agents are no longer as intelligent as they used to be. They frequently fail to locate files, even when I have already @-referenced the file! When this happens, the system automatically creates the file in an unexpected directory and writes partial content into it. Unfortunately, this content is incomplete and has no practical effect. Moreover, the system executes unnecessary instructions, such as creating folders that I do not need or directly compiling my program, while all I want is to update the @-referenced file. The constant inability to locate files and the execution of irrelevant actions make my project chaotic.
Issue Two: Auto Mode Selection and Peak Period Problems
I have noticed an interesting issue regarding the auto mode selection model and the slow user peak period. The peak period usually occurs at night, around 3 o’clock New York time. During this time, Pro users who are on the slower subscription tier often encounter situations where neither Model 3.5 nor Model 3.7 can be used, or only one of them is available.
I checked the auto-selection option, which defaulted to Model 3.5 to provide services for me. However, it prompted me that there was a peak period and suggested I use Model 3.7 instead. I manually switched to Model 3.7 and found that it was functioning properly! This indicates that the auto-selection feature defaults exclusively to Model 3.5, which appears to be ineffective during peak periods.
Another frustrating aspect is that when slow users enter the peak period, I may have already started a conversation. There might not be a peak period at the beginning of the conversation, and the system might have already replied to me. However, midway through the operation, another peak period occurs, requiring me to resend the request. While I understand the necessity of peak period restrictions, I believe the system should wait until the current reply is completed before enforcing these restrictions, rather than interrupting the process halfway through.
Issue Three: Decline in Intelligence and Efficiency
Cursor has become significantly less intelligent than before, and it no longer follows my instructions as accurately as it once did. Before version 0.46, it usually took only three or four requests to complete a task. Now, however, it often requires more than ten requests. Initially, I thought this was because the Cursor team had downgraded our model to ensure better service quality for faster-tier users. To address this, I subscribed to Cursor again within the same month. Unfortunately, the situation remains unchanged. Cursor is still not following my instructions properly.
Additionally, the official version of Model 3.7 includes data for 2025, but Cursor only provides us with data for 2023. Although this detail may not be directly relevant, it reflects how Cursor has become much less capable than it was before version 0.46.
Final Thoughts
I hope the Cursor team will focus on resolving these existing issues instead of repeatedly shifting responsibility to Claude. I love Cursor very much because it has significantly reduced my workload in the past. However, the current state of the tool has left me deeply disappointed. I am now considering whether I should look for alternatives.
