Build an AI Board of Advisors in Claude Code With a Free Skill
How the free "Roast Room" Claude Code skill simulates a debate among AI personas like Bezos and Buffett to help you make decisions.
What is the “Roast Room” Claude Code skill?
Roast Room is a free Claude Code skill that simulates a debate among AI personas modeled on well-known business figures, including Jeff Bezos, Warren Buffett, Paul Graham, Alex Hormozi, Gary Vaynerchuk, Dan Martell, and Gino Wickman. You give the group a decision with multiple options, and each persona scores the options, argues its position, and eventually delivers a joint verdict. It runs inside Claude Code as a skill invoked with a slash command, and every persona is built as a separate sub agent with its own “brain” of scraped source material.
TL;DR
- Roast Room is a Claude Code skill that runs a multi-round debate between AI personas based on real public figures, then returns a scored verdict on a decision you submit.
- Each persona is a separate sub agent, not a single chatbot playing different characters, which the creator says keeps answers more distinct and keeps costs lower than running full agent teams.
- Personas are built from persona “brain” files, markdown documents covering frameworks, decision rules, past stories, voice, and sourced material scraped from each person’s public content (one example cited around 124 sources and 1.3 million words).
- The debate runs in three rounds: opening positions and scores, cross-examination where personas argue with or against each other, and a final vote that produces a written verdict.
- A Claude Code chair (the facilitator role baked into the skill) decides which personas need to weigh in and can run a debate with only a subset of the council if the question doesn’t require everyone.
- The skill generates a downloadable PDF report of the scores, arguments, and final recommendation, and stores past sessions so you can review or replay them later.
- An optional visual interface shows the personas “talking” in a roast room style layout, but it’s cosmetic. The underlying skill works the same without it, straight from the Claude Code terminal.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
How does the sub agent architecture work?
Each persona in the council is its own Claude Code sub agent, defined by a dedicated agent file. That file sets up the persona’s identity, the one core question it filters every decision through, and pointers to a larger set of reference files (its “brain”) that the sub agent reads when it needs more depth.
The creator specifically built this on sub agents rather than full agent teams after testing both. According to the video, agent teams and sub agents produced similar quality answers, but sub agents were cheaper to run and preserved each persona’s distinct voice more reliably. A single “chair” role, played by Claude Code itself, facilitates the whole process: it assigns which personas argue for which position, flags when the group converges too quickly, and manages the three rounds that make up a session.
This structure means the council isn’t one model roleplaying seven characters in a single context. It’s seven independent sub agents, each reasoning from its own file, coordinated by a central facilitator. That separation is what lets each persona “own” a specific lens on the decision, like Bezos’s focus on reversibility or Buffett’s focus on cash risk, without those perspectives bleeding into each other.
What’s actually inside a persona’s “brain”?
Each advisor has a brain made up of multiple markdown files rather than a single personality prompt. Based on the Hormozi and Bezos examples shown in the video, a typical brain includes:
- A “hot sheet” summarizing the person’s worldview and roughly eight rules they decide by.
- Named frameworks associated with that person (for Hormozi, things like the value equation and grand slam offer).
- Documented past positions and decisions.
- Stories the person has told publicly, used to make responses feel more personal and less generic.
- A voice guide for how they phrase things.
- A sources file listing where all of this was pulled from.
The creator says this material came from scraping public content: YouTube videos, podcast transcripts, blog posts, and LinkedIn or X posts attributed to each figure. For the Hormozi persona specifically, the video cites around 124 sources and roughly 1.3 million words of source material distilled into the brain files. The same process was repeated for each of the seven personas.
This is also what makes the council customizable. Because each persona is just a folder of markdown files plus an agent definition, you can ask Claude to build a new persona the same way, pointing it at a different public figure’s content, and swap that person into or out of your council depending on your industry or the type of decision you’re making.
How does a Roast Room debate actually run?
A session starts when you invoke the skill (shown in the video as a slash command) and submit a decision brief: the question, the available options, relevant business context, available resources, and what success would look like.
From there, the debate runs in three rounds:
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
Round one. Each advisor sub agent writes an opening position independently, picks an option, and scores it, typically on a 1-to-10 scale, with reasoning tied to their specific lens (Bezos on reversibility, Graham on growth and demand, Buffett on cash risk and long-term value, and so on).
Round two. The chair reviews round one. If the group converges too quickly on one option, the chair can assign specific personas to argue hard for a different option to stress-test the consensus. Personas respond to each other directly, agreeing, disagreeing, or pushing back on specific points.
Round three. Each advisor casts a final vote. The chair compiles everything into a verdict: the recommended option, the conditions attached to it, and the key points raised along the way.
In the example walked through in the video, a conference pricing decision with four options, the council converged on one option with an average score of 7.1 out of 10 in round one, debated it in round two, and confirmed it in round three with specific conditions attached (a price point, a capacity cap, and a go/no-go threshold before committing).
Is the visual interface necessary to use it?
No. The visual layer that shows personas as animated characters “talking” in a roast room is a cosmetic add-on built on top of the skill. Everything it displays, opening positions, round-by-round arguments, and the final verdict, is already being generated and saved by Claude Code itself. You can run the skill purely from the terminal and get the same PDF report and session history without ever opening the visual interface.
The interface does add convenience: it lets you watch the debate happen live, browse past sessions, and replay a previous council discussion from saved data. But the underlying mechanics, the sub agents, the chair, the brain files, and the verdict generation, are all part of the free skill itself, independent of any front end.
Is it worth setting up for your own decisions?
The honest answer depends on what kind of decisions you’re making. The format is suited to choices with a handful of discrete options and real tradeoffs, pricing tiers, hiring versus not hiring, which market to prioritize, rather than open-ended brainstorming. Because the chair can detect when a question only needs input from a subset of the council (the video gives an example where only four of the seven personas were relevant to a hiring decision), it scales down automatically rather than forcing every persona into every debate.
The creator is clear that the output isn’t meant to be followed blindly. The value is in seeing a decision argued from several distinct, well-documented points of view, which can surface considerations you hadn’t weighed, even when you ultimately disagree with the final verdict. For anyone who already uses Claude Code regularly, the setup cost is mostly in cloning the repo and letting Claude install the agent and brain files into a project, after which running a session is just submitting a brief through the skill.
Frequently Asked Questions
What is a Claude Code skill?
A skill in Claude Code is a packaged capability, often a slash command, that bundles instructions, agent definitions, and reference files so Claude can perform a structured, repeatable task rather than relying on a single freeform prompt.
One coffee. One working app.
You bring the idea. Remy manages the project.
Do I need to pay for this to use it?
The Roast Room skill itself is free and distributed as a repository you clone into a Claude Code project. Usage still runs through your existing Claude Code access, so normal API or subscription costs for running Claude apply.
Can I change who is on the council?
Yes. Because each persona is built from a sub agent file plus a set of brain files sourced from that person’s public content, you can ask Claude to research and build a new persona in the same format and swap it into the council in place of an existing one.
Why use sub agents instead of one model playing all the roles?
The creator tested both sub agents and full agent teams and found the answers similar in quality, but sub agents were cheaper to run and did a better job preserving each persona’s distinct voice and reasoning style across a debate.
What do I get at the end of a session?
A final verdict with the recommended option, any conditions attached to it, and a summary of key arguments from each round, delivered as a downloadable PDF. Past sessions are also saved so you can revisit or replay them later.
