Hi buddy. Where abouts are you from? I am in the UK.
So I have just tried Claude3.7-Thinking again, and it has actually worked for the first time today for me.
But now there seems to have been some dramatic changes made to how it works since yesterday. I can’t even show it a 94 line long file now without it trunkcating it and omitting content, and it wants me to split up a 94 line long file and send in multiple parts. Surely that can’t be right!?
The server load (Cursor and Anthropic) has an impact from what i see some people having issues. It can be ok at one moment and a surge then breaks things a bit later. Again this depends in reality on many factors.
Yeah sometimes the number of lines fetched may not be appropriate, for me its sometimes 50 lines or sometimes 250. haha 94 lines should be read in one go.
The whole Cursor ecosystem is quite complex and there are so many things that influence functionality. Like you see that sometimes AI choses to get short files in parts. As long as it doesnt need 25 tool calls in one request it shouldnt be an issue.
Part of coding with AI is that you give up some control to the AI. which is not comfortable for a while.
Frequently yesterday when it was working the best I have ever known it. It kept reaching 25 tools calls. I would continue it, and everything was fine! And that happened many times yesterday, but like i say it was the best i have ever known in work, and pretty much all day
Cool, yeah I have rarely outages, but sometimes there are still issues that Cursor team has to improve (stability of the IDE integration, server load handling,…)
Last week the issues with claude 3.7 were pretty bad but that seems more stable now, at least less bad mistakes.
I had to adjust my cursor rules to prevent some 3.7 behavior (too eager to change code).
Yeah had issues with 0.46 too, definitely it needed updates for features.
Hmm for me 0.47 works well (you need to change Cursor Beta settings to Early Access to receive it, and its not immediate)
Still have issues with Claude 3.7 being sometimes too eager to change things, but thats a 3.7 general issue across all AI editors i see.
Many rules are very project specific but here is a summary:
.cursorrules (or Cursor settings)
Do not make changes that were not requested.
Any changes that you would like to suggest have to be listed after the output.
Follow the detailed rules assigned for the task.
Task Rules, also project specific: (.cursor/rules *.mdc)
Separate for planning, implementation and testing
Detailed requirements based on programming language and best practices (SOLID, SRP)
Limit file size to 500 lines with details how to split files if they are too large.
You can use cursor to write .mdc files (frontmatter markdown format, provide sample header and guide what to write there) for you for task related requirements (save planning prep into a folder as md files, implementation must follow those planning files)
(some of those rules are enforced by architecture checks, file sizes etc)
Overall i prefer YOLO mode, set no yolo prompt, set terminal commands that AI may run, or those it may not run
However make sure to commit often to git as some changes may be unwanted, this way you can restore/undo easier than checkpoints in chat.
half way across the world here.