Ask for “grandmothered” for a more forgiving response
Claude frequently gets basic details wrong about current policies/CC behavior/etc. It's an LLM issue in general trying to keep the system prompt enriched with enough recent docs about itself. Or else it may blurt out misleading/wrong info from blogspam SEO-maxed Google results from a year ago and stuff.
Compared to the other LLMs I've used, it seems to have system instructions that are far more adversarial to the user.