I feel like rules have been worse and worse. The option doesn’t even appear in “@“ anymore, I need to go to parameters to manage my rules.
But, in reality, it’s pretty useless since I feel like the agents won’t use them at all. No wonder if I’ve set up the rule to always be applied or conditionally. The only time it works is when I reference them manually like a classic file. But in this case, it’s not a “cursor rules“, it’s just a .md with instructions I can reference… not as useful.
It would be great to see them working again. My co-worker are using Claude Code and the CLAUDE.md (Claude alternative to rules) is a central part of using Claude code (and BTW : I activated the option to use CLAUDE.md with Cursor, but that doesn’t work either).
Which models do you use? I refrain from using cheap ■■■ models as they are worthless and waste time. Please be sure to stick with the smartest ones. Here are my top preferences:
Claude Sonnet/Opus 4.5 thinking
GPT-5.1 Codex Max XHigh Fast
Composer
I’ve never encountered these models not following the rules. Make sure your rules are short and to the point. I’ve seen people writing long ■■■ prompts in their rules that do not play well with LLMs.
I’m always using Claude Sonnet / Opus 4.5 or composer. I think I saw an article from Anthropic that proposed to tell the Agent to call you by a special name to be sure you instructions is taken into consideration. I’ll try that.
