Image: GitHub
Why it mattersA founder can get seven cross-domain reads on a hard call without hiring an advisor, and every quote in the verdict is checked against a sourced brain so the thinking is auditable later.
A founder with a hard call rarely has seven people to run it past. Roast Room, an open-source Claude Code skill, puts seven simulated advisors around a table and argues the decision for the founder. The repository is at 622 stars on GitHub.
The council has one seat each for Jeff Bezos on strategy, Paul Graham on product, Alex Hormozi on offers and pricing, Gary Vaynerchuk on marketing, Dan Martell on hiring and founder time, Gino Wickman on operations, and Warren Buffett on money and capital. Each seat is a Claude Code sub-agent that reads its own brain folder of frameworks, principles, positions and quoted lines, each tied to the public talk or essay it came from. A quote-check script fails a council when a line in the verdict is not in a brain with a source.
How a session runs
A user types /roast-room followed by the decision, the options, the numbers (revenue, cash, team size), the limits (deadline, hours, budget), what success looks like and the way the founder is leaning. The Chair, played by Claude Code, asks for anything missing, then runs the rounds. Each seat scores every option, weighted by how much that advisor's pillar matters to the call in front of them. A full council is twenty-one sub-agent runs. A --quick flag seats four advisors for two rounds and uses far less of a Claude plan, for smaller and reversible calls. A --seats flag picks named advisors, and ask gets one answer from one seat.
A small Node web app runs on http://localhost:4730, renders the council around a table and plays the Chair's steps back as it writes them. The web app makes no model calls. Three example sessions ship with the repository so a new user can watch one before running their own. Everything saves to sessions/<date>-<decision>/verdict.md plus a one-click PDF.
The honesty of the simulation
The repository states out loud that these seats are not the people and do not speak for them. The brains only quote from each advisor's public work, and sources.md next to each brain lists what the brain was built from. Two further rules guard the verdict. The Chair caps a verdict when any seat flags a ruin risk, and it forces a real argument when every seat agrees too fast. The second rule exists because the point of calling in an advisor is to be challenged, and one who always agrees does not challenge.
The skill is MIT-licensed and needs Claude Code plus Node.js 20. There is no backend, no account, no telemetry. The Claude plan covers model calls, so a session costs no cash out of pocket, only usage on the plan. An optional runner-mode variable bills an Anthropic API key instead, at about 25 calls per full council.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.