When I tell Claude-4-sonnet about an error log, he always adds try…catch everywhere in the file to avoid certain possible errors. However, this is just a way to cover up the problem, and it doesn’t really solve anything. Because in fact, it’s just an object reference error.
I try to add: Do not use try…catch to fix any problem in Cursor Rule. However, it did not help. It still ignored my rule.
Expected Behavior
It should be smarter
Operating System
MacOS
Current Cursor Version (Menu → About Cursor → Copy)
If you have a problem within the tool calls - i.e. text generation coming deliberately from the chat model - you have two points of possibility here:
-The model is either trained outside, not properly aligned with, or not sufficiently overlapping with code examples that reflect your project versioning, overall app architecture, and then also we have to really stay on top of it inserting slightly incorrect variable names, etc. This is Claude-4-Sonnet’s culpability. HOWEVER, this is where it gets complicated and i empathize with you.
These models, when prompted in sequence is still just IO with clear consequences - and even though in modeling we do not understand true Error, we still attempt to quantify and explain it in or out of outputs (we are programmers so this is easily automated with a local LLM, a 3080/4090, it’s just not sufficient hardware power and you run into more fluffy incorrect answers, or ones that are SO close to the task, but therein lies the problem - the model is predicting the next word, not solving your problem.
What is solving your problem is altering and taking the outputs of the embeddings given your query and what we know and don’t about the offered model, is what Cursor works hard at and generally does well lately - using chains, prompts, embeddings, and tool calling systems to manifest this model’s output task indirectly to write you code.
The reason you think this specific behavior is Cursor’s fault, and why you can’t do anything but say “it should be smarter”, is not your fault. It is actually the correct conclusion to make - it should do X - there is expectation there. And these things fluctuate in handling, and therefore cost, and instead of being explained or investing in understanding the companies they partner with to offer these models, they act like it’s all ONE BIG THING and use terms like cost insensitive to condition users into thinking unsure but usage based costs are “OK” if i set a hard 50 dollar limit, and say in tiny font “cost insensitive” which is what that solution is. Which I find extremely concerning in the scope of what I know to be the arrangement here and also what I know about LLMs, and what I’ve raised as issues on these boards about countless times, regardless of tone. It is a problem.