It's a language model. Language is what it models. So you use language to move it into an advantageous state space. Dunno what you want from me, lol.
Ultimately you're not conversing with any real LLM, it's iterative document generation where humans perceive fictional characters in the output. If the text you contribute says "You're just an {Noun}" that's that's shaping the document output based on what got trained in relation to {Noun}.
Which may eventually backfire, when the (real) LLM gets trained on documents such as blog-posts like "The moment you realize the {Noun} is retarded."