Seems less like an "alter ego" and more like getting it to nest a response within another response, which implies that the language model generates responses in multiple passes and the filters are only applied up to a certain depth