---
title: How to Build an AI Expert Council in Claude Code (Roast Room)
url: "https://goremy.ai/blog/articles/roast-room-claude-code-ai-expert-council"
author: Luis Chavez-Mattos
editor: Luis Chavez-Mattos
editor_title: Director, Product
published: "2026-10-30T09:00:08.122Z"
updated: "2026-10-30T09:00:08.122Z"
topic: AI Agents & Automation
description: The Roast Room is a council of persona sub-agents in Claude Code that debate your idea and vote. See how it works, how to add members, and why sub-agents.
---

# How to Build an AI Expert Council in Claude Code (Roast Room)

> The Roast Room is a council of persona sub-agents in Claude Code that debate your idea and vote. See how it works, how to add members, and why sub-agents.

Edited by Luis Chavez-Mattos, Director, Product · Published 2026-10-30

## What is the Roast Room in Claude Code?

The Roast Room is a council of AI personas that pressure-tests a business idea before you act on it. Each persona is a Claude Code sub-agent modeled on a public figure, such as Alex Hormozi, Jeff Bezos, or Warren Buffett. You submit a decision, each member states an opening position, they respond to one another, then they vote and the main agent summarizes. Builder Nate Herk demonstrated the setup on a recent guest episode of a workflow series.

## TL;DR

- The Roast Room is a **council of persona sub-agents** in Claude Code, each built from research on a real public figure.
- A session runs in stages: **opening positions, debate, a vote, and a summary** from the main agent.
- A **new-roster skill** handles adding members, so you do not hand-write each persona file.
- The design rests on a simple principle: **start from proven work** and break down how successful people reached their results.
- **Sub-agents** fit this job because each persona only needs to report back to a coordinator, not coordinate with peers.
- Treat the output as **structured pushback**, not expert advice. The personas are approximations built from public material.

## Why build a council instead of asking Claude once?

A single prompt gives you one blended answer. Language models predict likely text, so left alone they drift toward what most people would assume or what sounds sensible. That is the opposite of what you want when you are stress-testing a plan.

The Roast Room source makes a related point about working with AI in general. The best inspiration is proven work. In production AI you test against a golden dataset, so it makes sense to start with one when you build. For a business decision, that means finding people who did what you want to do, then breaking down how they got there.

A council turns that idea into a repeatable tool. Each persona carries a distinct set of priorities, so disagreement shows up on the page instead of getting averaged away.

## How does the Roast Room work?

The flow has four stages. The main Claude Code session acts as moderator.

1. **Brief.** You describe the idea, the numbers you have, and the decision you face.
2. **Opening positions.** The moderator calls every persona sub-agent with the same brief. Each returns an independent take. Because they run separately, nobody anchors on anyone else's answer.
3. **Debate.** The moderator passes the opening positions back to each persona and asks for a rebuttal or a revised stance.
4. **Vote and summary.** Each member votes (for example proceed, change, or kill). The moderator tallies the votes and writes a summary covering the main objections, where members disagreed, and what to test next.

One practical detail: sub-agents report to the session that spawned them. They do not message each other. So the debate happens in rounds, with the moderator relaying positions. That is not a limitation in practice, since it keeps the transcript in one place and lets you control how many rounds run.

## How do you build the persona sub-agents?

In Claude Code, a sub-agent is a markdown file with frontmatter, typically stored in `.claude/agents/`. Each runs in its own context window with its own instructions.

```markdown
---
name: persona-buffett
description: Council member who evaluates ideas through durable advantage, price paid, and downside risk.
tools: Read, WebSearch
---

You are a council member modeled on Warren Buffett, based on his
public letters, interviews, and stated principles.

When given a business idea:
- State your opening position in under 200 words.
- Name the single biggest risk you see.
- End with a vote: PROCEED, REVISE, or KILL.

Stay in character. Do not claim to be the real person.
```

The quality of a persona depends on the research behind it. A thin prompt that says "act like Bezos" produces a caricature. A better file grounds the persona in what the person has actually said and done: books, shareholder letters, long interviews, documented decisions. Pull those into the file as principles, typical questions, and known blind spots.

Give each persona a distinct lens. A council of five members who all love growth is just one opinion repeated. Mix an operator, an investor, a skeptic, and someone who cares about the customer.

## How do you add new members with a roster skill?

Writing each persona by hand gets old fast. A skill can automate it. Skills in Claude Code live in a folder with a `SKILL.md` file that describes when and how to run a repeatable procedure.

A new-roster skill would do roughly this:

1. Take a name and, optionally, what you want from that person (pricing, hiring, risk).
2. Research public material about them, using web search or documents you supply.
3. Extract principles, decision habits, and typical objections.
4. Write a new agent file in the same format as the existing members.
5. Add the member to the roster the moderator reads from.

Keep the roster in one file the moderator consults, so adding someone never requires editing the debate logic. Review each generated persona before trusting it. Check that the claims trace back to something the person actually said.

## Why sub-agents instead of an agent team?

Claude Code offers more than one way to run several agents at once. The two relevant options are sub-agents and agent teams. Agent teams, an experimental feature, let multiple Claude instances work in parallel and communicate with each other directly. Sub-agents work for a single coordinating session and send results back to it.

For a council, sub-agents are the natural match:

| Factor | Sub-agents | Agent team |
|---|---|---|
| Communication | Report to the moderator only | Teammates can message each other |
| Structure | Fixed stages you control | More open-ended collaboration |
| Cost | Generally lower, short focused runs | Generally higher, each member is a full session |
| Fit for a vote | Moderator tallies and summarizes | Needs extra coordination to converge |

A roast has a predictable shape: collect opinions, exchange them, vote, summarize. You do not need members negotiating freely. You need independent takes and a single place where they get compared. That favors the simpler, cheaper tool. If your use case involved members building something together over a long task, a team would make more sense.

## Is a persona council worth using?

It helps most as a cheap way to find holes. You get objections you had not considered, framed in distinct voices, in a few minutes. It works less well as a decision-maker.

Keep these limits in mind:

- Personas are built from public material and the model's interpretation of it. They are not the real people.
- Models tend to be agreeable. Tell personas explicitly to disagree where warranted and to name the strongest reason to say no.
- A vote of 4 to 1 is not evidence. It is a prompt to read the dissent.
- Your own numbers matter more than any persona. Feed the council real data.

## Frequently Asked Questions

### Do I need Claude Code agent teams to build this?

No. The council works with ordinary sub-agents and a moderating main session. Agent teams are an option if you want members to message each other directly, but a roast only needs independent opinions and a moderator.

### How many personas should the council have?

Enough to cover distinct viewpoints, and no more. Between three and six is a sensible range. Every extra member adds cost and lengthens the summary, and similar personas mostly repeat each other.

### Can sub-agents debate each other directly?

Not on their own. A sub-agent reports back to the session that called it. To create a debate, the moderator shares each opening position with the others and asks for responses in another round.

### Are the personas accurate representations of the real people?

They are approximations. A persona reflects the public material you use and how the model interprets it. Use them as lenses for pressure-testing, and verify any specific claim attributed to a real person.

### Can I use this for things other than business ideas?

Yes. The same structure suits pricing decisions, product positioning, hiring plans, or content strategy. Change the roster to match the domain, and keep the staged flow of positions, debate, vote, and summary.