Team GPT: Simulating Multi-agent Collaboration With Role-playing Prompts

Have you ever been to a great meeting? The right people with a diverse set of skills, knowledge, and perspectives, where all the inputs are synthesized into a single, coherent output that addresses the issue being discussed.

I have (twice!) and it’s a marvelous thing. The only problem is that this involves humans, which are expensive and messy. So it seems reasonable to replace them with something more reliable that doesn’t need as much water.

Below I’ll share a few experiments in getting GPT to play the role of multiple participants in a discussion, incorporating inputs from characters with varying thinking styles, and combining these inputs into a single output.

The goal is that when facing any new, complex, multi-faceted problem (particularly problems that impact many lives), you can employ such a prompt to view the issue from several angles, resulting in a well-rounded opinion or solution.

The prompt

There are three parts to the prompt. First, a description of the setting and the cast of characters. This can be reused with little modification for different problems.

A group of colleagues have assembled to discuss a particular issue.

William is neutral and objective, concerned with data, facts, figures, and information. He likes to be clear about what is known and unknown, and what additional data needs to be gathered.