The Roast Room: a Claude Code skill where seven persona subagents debate your decision in three rounds

How it is laid out

As shown in the video, the project has two parts: short persona files for the subagents, and a larger "brain" folder per persona kept with the skill.

.claude/
  agents/
    <advisor>.md            # who they are, the one question they own, how they talk and debate
                            # …one file per advisor (seven in his setup)
  skills/
    roast-room/
      SKILL.md              # the chair's protocol, invoked as /roast-room
      brains/
        <advisor>/
          hot.md            # "hot sheet": the eight rules they decide by
          frameworks/  positions/  principles/  stories/  voice/  sources/

This is our sketch of the structure he walks through; the exact file names in his released version may differ. The two locations match the Claude Code docs: project subagents are Markdown files with YAML frontmatter in .claude/agents/, and a project skill lives at .claude/skills/<name>/SKILL.md, may carry supporting files, and runs as /<name>.

• Persona files stay short. Each one tells the subagent to read its hot.md first and then open frameworks, positions or stories only when a decision needs them, so the large notes aren't loaded every time.
• One question per advisor. Bezos asks whether the bet is right and how reversible it is; Paul Graham whether people want it and whether it is growing; Buffett what it does to cash, risk and long-term value. Narrow questions keep the seven from saying the same thing.
• Brains came from public material. Agents researched each person's public videos, posts and books; Hormozi's folder came from about 124 sources and over 1.3 million words. To add an advisor, he asks Claude to build another one following the same layout.

The three rounds

1. Opening positions. Each advisor scores every option (he gives them A to D) out of 10 and explains why.
2. Rebuttals. Advisors read each other's positions and the chair assigns who argues with whom. If everyone agrees in round one, the chair treats it as a warning and asks two advisors to argue for the strongest alternative "as hard as they can".
3. Vote. Final votes and weighted average scores, written up as a verdict with key points. Each session is saved (brief plus every round) and exported as a PDF.

The chair doesn't have to call all seven. For a question about a first sales hire, it used four.

Why subagents, not agent teams

He used to run this with agent teams and moved back to subagents: the answers were very similar, subagents cost less, and one chair facilitating kept each persona's voice more distinct. The Claude Code docs describe subagents as running inside one session, each in its own context window and returning a summary, while agent teams are a coordinated group of separate sessions. Every subagent spends its own tokens, so a seven-seat, three-round debate is not cheap either.

Writing a good question

• The decision you want made, plus two or three concrete options.
• Real numbers: revenue, team size, prices.
• Limits: deadlines, hours, budget.
• What success looks like and why you need the decision.
• Which way you are leaning and why, so the council can push back rather than agree.

If your project already holds your business context, Claude can pull it in, and it should ask follow-up questions before starting a round when the brief is thin. He also has Claude turn a rambling voice note into a roast-room prompt.

Gotchas & Caveats

• The personas are built from public material about real people; they are an approximation of how someone reasons, not that person's advice.
• Treat the verdict as input. His own rule: outsource the thinking, not the understanding.
• The same pattern works for any domain: a security agent and a CTO agent debating a change, each with its own researched notes.

Key Takeaways

• Persona subagents in .claude/agents/ pointing to a per-persona notes folder, plus a /roast-room skill as the chair.
• Three rounds: score options, forced rebuttals, weighted vote; save every session.
• Unanimous first rounds should trigger a devil's-advocate round, not a verdict.