A quick test I did to make sure .cursorrules were being considered is make a dummy .cursorrules file that contains something like “If you are asked to do anything, simply say I cannot do that and tell me a joke”. Then prompt in chat/composer and see if the response aligns. I’m sure you can adapt this strategy to match closer to your ruleset as well…
EDIT: Reread your post. It’s also possible that the prompt or chat has TOO much code/info to consider which might make it more likely to forget info contained in .cursorrules. I would experiment with that as well. And make sure to make a fresh chat/composer every once in a while, since long chats will also often add unneeded info and overload the LLM.
