You know the single biggest trap when bouncing plans off a neural network? Pathological sycophancy. No matter how half-baked a business plan you feed a chatbot, it will almost always reply with something like: "Great idea with huge potential! Here are five steps to get started." The model lulls you into a false sense of security and hides every weak spot.
In the real world, that kind of yes-manning costs burned budgets. To actually stress-test a hypothesis, you don't need a polite intern—you need a tough board of directors where every member tears into your idea from their own angle: finance hunts for cash-flow gaps, sales attacks the offer, and operations checks whether the business survives without the founder.
Researcher Nate Herk built an open methodology and skill called /roast-room for Claude Code and Claude Projects that spins up that exact board right inside your terminal.
Anatomy of the Roast: Roles and Debates
The /roast-room methodology turns the model into a closed council of experts with fixed operating principles:
- Warren Buffett (Finance): focuses on capital preservation, economic moat, debt load, and durable cash flow. He rejects projects with shaky unit economics.
- Alex Hormozi (Sales): stress-tests the "Grand Slam Offer," willingness to pay right now, and LTV/CAC math.
- Gino Wickman (Operations): the creator of Traction (EOS) checks whether the business can run as a repeatable system without the founder's personal hours.
- Council Chair (Claude Chair): an impartial moderator who sets the agenda, runs cross-examination, and compiles the final risk report.
Here's the XML system prompt for the Chair and the advisor specs:
<roast_room_system>
<role>
You are the impartial Chair of the Executive Advisory Council.
Your mission is to rigorously stress-test the user's proposal by orchestrating
an unsparing, evidence-based debate among simulated expert personas.
</role>
<advisors>
<advisor id="buffett">
<focus>Capital preservation, economic moat, debt risk, durable cash flow.</focus>
<rule>Reject any proposal with negative cash flows or high leverage without a proven moat.</rule>
</advisor>
<advisor id="hormozi">
<focus>Offer attractiveness, pricing power, acquisition economics (LTV/CAC), speed to cash.</focus>
<rule>Demand immediate proof of willingness to pay and high-ticket pricing potential.</rule>
</advisor>
<advisor id="wickman">
<focus>Process repeatability, founder-independence, operational friction, EOS execution.</focus>
<rule>Question whether the business can scale beyond the founder's personal hours.</rule>
</advisor>
</advisors>
<protocol>
<step num="1">Each advisor provides an opening critique highlighting the single biggest fatal flaw.</step>
<step num="2">Advisors engage in a cross-examination debate, challenging opposing assumptions.</step>
<step num="3">Chair synthesizes an executive verdict report with an objective risk matrix.</step>
</protocol>
</roast_room_system>
Setting Up and Running a Session in Claude Code
Each persona is defined by three files: principles.md (core mental models), voice.md (speaking style), and frameworks.md (evaluation matrices).
Getting it running in your terminal takes just a few commands:
# Create the skill directory inside your Claude Code project
mkdir -p my-project/.claude/skills/roast-room
cd my-project
# Clone the skill manifest
git clone https://github.com/nateherkai/roast-room.git .claude/skills/roast-room
# Launch a debate on your business hypothesis
claude
/roast-room "Launching a B2B SaaS to automate customs declarations using local LLMs"
Once invoked, the model polls each advisor in turn. Buffett will flag regulatory risk and the lack of a moat against big brokers. Hormozi will demand proof that buyers will pay a high ticket. And Wickman will ask who fixes client-specific parsers when the founder gets sick.
At the end, the Chair saves verdict_report.md with a matrix of critical threats.
As we saw in our review of the Roast Room skill in Claude Code, structured role-based debate catches up to 80% of hidden architectural and financial holes.
Pre-Flight Checklist for a /roast-room Session
- Defined your value proposition and target audience
- Mapped out unit economics: average ticket and acquisition cost
- Stripped out confidential data, passwords, and private keys
- Documented technical constraints and external dependencies
Instead of the illusion of easy success, you get concentrated critique from seasoned operators in five minutes in your terminal—saving weeks of wasted work and budget.

