The actual question is backwards: how do we keep the prompt and context small enough so the LLM doesn't start hallucinating basic rules of chess.