I Built the Ultimate AI Board of Advisors (free claude skill)

Summary

Nate Herk demonstrates the Roast Room, an advisory board system running as a skill inside Claude Code. The system convenes simulated council personas representing Jeff Bezos, Paul Graham, Alex Hormozi, Gary Vaynerchuk, Dan Martell, Gino Wickman, and Warren Buffett. Each persona focuses on a distinct business lens. Bezos assesses whether a decision is reversible and a good bet, Graham evaluates genuine user demand, and Buffett analyzes cash impact, risk, and long-term value. Herk demonstrates the system by feeding it a founder dilemma regarding whether to commit to an in-person conference in Austin in June 2027 or pursue virtual alternatives.

The council deliberates across three rounds coordinated by an agent acting as the Chair. In round one, each advisor evaluates the brief independently and generates an opening position. This prevents the personas from anchoring on whoever speaks first. The advisors score options from one to ten, with the Austin demonstration showing unanimous initial support for a scaled-down in-person tier. When an option achieves universal backing, the Chair triggers an intentional counter-argument phase in round two, ordering specific personas like Buffett and Martell to push back and argue hard for alternative paths. Personas challenge one another directly using their core operational principles.

After cross-examination, round three concludes with final votes and a comprehensive verdict. For the demonstration query, the council returns an approval labeled Go with conditions, specifying that the organizer must pre-sell fifty paid memberships before signing venue contracts and establish hard refund deadlines. The skill stores all session output into markdown files, including event logs, intermediate rounds, and a final verdict document. Herk shows how this output is rendered in a custom visual web interface with avatar speech bubbles, as well as an exported PDF report complete with advice breakdowns and recommended action items.

Herk opens the project code repository to illustrate how the advisors are configured. Rather than using unstructured system prompts, each persona possesses a sub-agent definition file mapped to a knowledge directory called a brain. For example, the Hormozi persona draws from 124 public talks containing 1.3 million words, organized into files detailing eight decision rules, named frameworks, stances, and illustrative stories. The Bezos persona contains data from twenty-four shareholder letters and thirty-seven public interviews. Herk emphasizes that users can duplicate this structure to build advisors for any domain, such as technical security or executive leadership, by scraping relevant subject matter expertise into structured markdown documents.