"Steer agents by shaping their context narrative Through hands-on experimentation with persona-driven agents, the speaker concluded that the model is always implicitly performing a character based on what is in its context window. Being 'mindful of the context that you give it while you're operating the model in such a way where you're nudging it towards personas that give you the kind of results that you want' is what the speaker called 'a sort of powerful underutilized skill.' Rather than relying solely on direct instructions, practitioners can improve output by deliberately curating the surrounding narrative and reference material in the prompt. Source: Speaker, AI Research Presentation, OpenClaw Molt Book & Agent Sociality Studies"