Nowadays my rule is "the comment should contain enough about the why and the how so that the LLM generates the next few lines mostly correctly."