I mean, if advice like this worked, then why wouldn't Anthropic let the LLM say it, for instance?

Because it’s embarrassing, and probably nobody understands why this works, depending on such heuristics that can completely change in the next model is really bad…