During role-playing, assign a role to each agent using system prompts/messages. Compared to other typical conversational language models, the hint engineering of the new framework is limited to the initial stage of role-playing scenarios. In this framework, the indicator initially acts as the CEO, participating in interaction planning, while the assistant assumes the role of CPO, performing tasks and providing responses.
To achieve role specialization, the researchers used inception prompting, which proved to be effective in enabling agents to perform their role functions. Prompters and assistants contain many important details, such as assigned tasks and roles, communication protocols, termination criteria, and constraints designed to prevent undesirable behavior such as instruction redundancy, no message response, infinite loops, and so on.