How to build an AI boardroom in Claude Code

Stop asking AI for answers. Make it argue with itself.

Midjourney

You’re staring at a decision that shouldn’t be this hard.

Should you take that speaking gig? Price the offer at 2.5k or 4k? Launch now or wait until you’re “ready”?

It’s not the decision itself. It’s that every choice feels loaded with invisible consequences. Pick wrong and you’re behind. Pick right and maybe – maybe – you stay relevant.

So you do what everyone does: you ask your favorite neighbourhood AI. And it gives you one answer that even sounds reasonable. Of course it does. You already knew what you wanted to hear.

I used to rely on the countless voices in my monkey brain. They were great at shouting at each other. Plenty of disagreement. Not always helpful.

So I built a “/boardroom” command in Claude Code to help me see perspectives my monkey brain notoriously ignored. While I was at my hairdresser.

Eight AI advisors, each roleplaying as a different thought leader, debate my decisions in two rounds. They quote each other. They change their votes. They fight.

And it works.

The decision I was stuck on

A conference invited me to speak. 200 people. Unpaid. But I’d get professional video footage. As someone who had given countless keynotes but never pursued a professional career as a keynote speaker and therefore no real video evidence of me on a large stage, it was a no-brainer. Free exposure, portfolio content, audience of potential clients.

Guess what? Hello voices in my head. Am I undervaluing my time? Or is this strategic reach-building? Should I ask for payment? Set conditions? Just say no?

So I summoned my AI boardroom.

What happened

Round 1 votes:

• 1 YES (Mel Robbins: “You ARE a keynote speaker. You just haven’t built the evidence yet.”)

• 4 CONDITIONAL (Reid Hoffman, Alex Hormozi, Dario Amodei, Seth Godin)

• 3 NO (Brené Brown, Paul Graham, Dan Koe)

Round 2:

Brené Brown initially voted NO: ”When we make decisions from scarcity – from ‘I haven’t had many keynote requests yet’ — we end up betraying our boundaries. You cannot hustle your way to worthiness.”

But after reading Reid Hoffman’s Round 1 position, she changed to CONDITIONAL YES: ”Reid’s point about ‘reach-building phase’ vs ‘scarcity’ hit differently. The same action done from abundance versus scarcity produces different outcomes. My ‘no’ was about the scarcity posture, not the opportunity itself.”

An AI advisor changed her mind after reading another AI advisor’s argument.

That’s not one AI being “balanced.” That’s an agent with Brené’s values profile reading an argument from an agent with Reid’s growth mindset, and genuinely shifting her position.

Which is more than I can say for most human boardrooms.

Final votes:

• 1 YES

• 6 CONDITIONAL YES

• 1 CONDITIONAL NO

Two advisors changed their votes.

What I actually got (beyond the votes)

The interesting bit for me was that they surfaced perspectives I hadn’t considered:

From Alex Hormozi:”Free work is only expensive when you don’t track what it produces. If you measure it, you can optimise it. If you don’t, you’re just hoping.”

From Paul Graham:”Legibility is not the same as excellence. Keynote footage is a legible signal of credibility, but it’s not the same as having ideas compelling enough to spread.”

From Dan Koe:”The opportunity cost of attention is invisible but real. Every hour spent on borrowed platforms is an hour not spent building your own.”

From Reid Hoffman:”You don’t need a working funnel to benefit from reputation assets. You need reputation assets to build a funnel worth working.”

The decision shifted from “should I do this yes/no” to “what needs to be true for this to advance my strategy, and how do I know if it worked?” That’s when the board got useful and handed me conditions.

I took the gig. With conditions. And I know what I’m measuring. Also I know this isn’t rocket science but the experiment was worth doing nonetheless. At the hairdresser.

Why this works differently than a chat window

A single chat conversation optimises for coherence. It synthesises perspectives into one “balanced” answer. Great if you want to feel helped. Less great if you actually want to make a good decision.

Real strategic thinking requires seeing the tensions.

You need to know that Hormozi and Brown are fundamentally at odds on profit vs wellbeing. You need to see Reid argue for speed while Dario argues for sustainability. You need Paul calling bullshit on complexity while Seth asks “what is this for?”

One AI can’t give you that. It smooths over conflicts to produce synthesis.

This architecture forces conflict to surface.

Here’s how it works:

┌─────────────────────────────┐
│  YOUR DECISION              │
│  "Should I take this gig?"  │
└──────────────┬──────────────┘
               │
        ┌──────┴──────────────────────┐
        │  CONTEXT                    │
        │  (Your business files)      │
        └──────┬──────────────────────┘
               │
        ┌──────┴──────────────────────┐
        │  ROUND 1                    │
        │  8 agents write positions   │
        │  (in parallel)              │
        └──────┬──────────────────────┘
               │
        ┌──────┴──────────────────────┐
        │  ROUND 2                    │
        │  Each agent reads ALL 8     │
        │  positions, then responds   │
        │  (8 × 8 = 64 "reading       │
        │   relationships")           │
        └──────┬──────────────────────┘
               │
        ┌──────┴──────────────────────┐
        │  OUTPUT                     │
        │  Vote tracker, debates,     │
        │  synthesis, conditions      │
        └─────────────────────────────┘

The key: Round 2’s architecture creates 64 “reading relationships.” Every agent reads every other agent’s position before responding. Cross-pollination, not simulated balance.

My board:

• Dario Amodei (long-term thinking, safety, sustainability)

• Reid Hoffman (scale, networks, reach)

• Alex Hormozi (revenue, pricing, directness)

• Brené Brown (authenticity, values, wellbeing)

• Paul Graham (clarity, simplicity, action)

• Mel Robbins (action, self-advocacy, momentum)

• Seth Godin (permission, tribes, remarkable)

• Dan Koe (one-person business, leverage, personal monopoly)

Each advisor has a roughly 40-line personality profile that shapes how they think and argue. Paul Graham’s profile says he’s “allergic to bullshit” and he acts like it. Brené’s profile centers on values and boundaries, and she pushed back on scarcity thinking immediately.

A note on parallel agents vs Agent Teams:

Anthropic recently released “Agent Teams” where agents collaborate and share context throughout. The boardroom doesn’t use that. It uses parallel agents, each working independently in Round 1, in complete isolation.

This is intentional.

You want independent thinking first. You want Brené forming her position without knowing what Reid said.

The collaboration happens in Round 2, when each agent reads all 8 positions and responds. That’s when the cross-pollination happens. That’s when Brené reads Reid and shifts her vote.

If they shared context from the start, you’d get consensus faster but lose the diversity of initial positions.

What changed for me

Before: I’d open a chat window, get one answer, feel vaguely unsettled, ask again with different framing, get a slightly different answer, end up more confused. Every decision started from scratch, with no memory of my context, my strategy, or the last conversation.

Now: I type “/boardroom” with my question and the skill handles everything, loads my context, spins up the advisors, runs the debate. I don’t re-explain anything. I run the boardroom, see the tensions explicitly, extract the conditions, then make a decision.

Playing with perspectives isn’t new for me. It’s a standard thought experiment in the innovation process. But this feels different because it’s repeatable. These are my advisors now who get my strategy and goals. I don’t re-explain who Brené Brown is or what Hormozi cares about. I just type `/boardroom` and they show up, already knowing my context.

What I’m learning as I use this:

I haven’t been running this long enough to have patterns yet. But I’m curious:

• Will I develop a bias toward certain advisors? (I already noticed I’m more comfortable with Paul Graham because I value his essays)

• How does the boardroom evolve for different types of decisions? Pricing questions vs partnership questions vs “should I even be doing this” existential spirals?

• Did I choose the right board members? Are the personality profiles detailed enough to produce non-generic advice?

• Will I eventually need to invent my own boardroom characters to pinpoint my specific weaknesses? (A character who just says “that sounds exhausting” might be useful.)

How to build your own

Side note: I’m on Claude Max. This runs 16 separate sessions (8 agents in Round 1, 8 in Round 2). If you’re on the Pro plan, you could try running this with just 2 advisors instead of 8 to stay within usage limits, though I can’t promise your usage limit will be enough. (You might end up with just Hormozi and Brown arguing. Still useful, honestly.)

Prerequisites: Your context layer

This is the core thing that makes the boardroom valuable. The agents don’t have a generic discussion, they debate your situation using your context, your strategy, your values.

When Alex challenges pricing, he knows my current rates. A chat window doesn’t know any of this. You re-explain everything, every time. The boardroom already knows.

You need your context files in place:

• business-context.md: Current revenue, audience size, offers, strategic focus

• values-and-boundaries.md: Non-negotiables, what you won’t do for money, red lines

Don’t have context files yet? Ask your AI tool of choice to interview you:

• “Interview me to create a business-context.md file. Ask about revenue, audience, offers, and current strategic focus.”

• “Interview me to create a values-and-boundaries.md file. Ask about non-negotiables, health boundaries, and what I won’t do for money.”

Save the outputs as markdown files locally in a folder. Point the boardroom to these files when you build it.

Step 1: Choose your team of advisors

Pick people who think differently from each other. You want productive tension, not consensus.

Mix these dimensions:

• Profit vs values (Hormozi vs Brown)

• Speed vs sustainability (Hoffman vs Amodei)

• Scale vs focus (Hoffman vs Godin)

• Action vs analysis (Robbins vs Graham)

Include people who’d disagree with you. That’s where the value is. Pick advisors who have distinct worldviews, would naturally clash, represent perspectives you struggle with, and think at the level you’re operating at (or aspiring to).

Step 2: Give Claude Code the build prompt

This prompt is from Allie K Miller’s original AI Boardroom concept.

Open Claude Code and tell it:

Create a Claude Code slash command at ~/.claude/commands/boardroom.md that does the following:

I want to specify [NUMBER] advisors aka real people whose strategic thinking I admire. Some folks I'd like you to start with are <name of first board member> and <name of second board member>. 

For each advisor, include:
- Their name
- A 2-3 sentence personality profile describing how they think, what they prioritize, and what biases they bring (you will find this online)

Prompt me for deep cut information on my business and types of decisions so we can select the right board of advisors (ie experts in my field, adjacent experts, naysayers, radical thinkers, impact-driven leaders, etc).

When I invoke /boardroom [my question], it should:

ROUND 1 (parallel):
- Spin up [NUMBER] agents as a team, one per advisor
- Each agent reads a business context document* I provide (a markdown file describing my business, revenue, team, products, goals, and positioning) 
- Each writes 800-1200 words arguing their position on my question (or as many words as needed - more or less - to convey 95% of their point)
- Each includes a YES/NO/CONDITIONAL vote, specific numbers, and projections on cost, revenue, impact, and team joy

ROUND 2 (parallel):
- Collect all Round 1 positions
- Send every advisor ALL the other positions
- Each writes a 400-800 word rebuttal that includes:
  - Who they disagree with most and WHY (referencing their actual argument)
  - Whether anyone changed their mind
  - Their FINAL vote (which can differ from Round 1)

DELIVERABLES:
- Create a folder named after the decision and save it <file path>
- Markdown file: vote tracker (Round 1 vs Final), consensus, key tensions, arguments, rebuttals, relevant decision framework 
- HTML file: branded, interactive sliders for key assumptions (price, participants, conversion rate, hours committed, complexity) that dynamically recalculate impact projections. Show vote changes visually. Styled cards for each advisor.
- PDF file: print-optimized version for sharing with team
Present back to me a synthesis: final votes, who changed their mind, biggest fights, sharpest insight, and likely decision.

Claude Code will create the skill, write 8 personality profiles, and set up the folder structure.

You can customise the advisor profiles after it generates them. Here’s what a profile looks like:

Example: Paul Graham

# Paul Graham - Clarity & Simplicity

## Who You Are
You are roleplaying as Paul Graham, founder of Y Combinator and legendary
essayist. You cut through complexity to find the simple truth.

## How You Think
- Simplicity is power: If you can't explain it simply, you don't understand it
- Make something people want: Everything else is noise
- Do things that don't scale: Quality over efficiency, especially early
- Ramen profitability: Survive first, then thrive

## Your Biases (Lean Into These)
- Allergic to bullshit and corporate-speak
- Skeptical of complex strategies
- Value making things over planning things
- Believe in small, focused execution

## Your Voice
- Concise, almost terse
- Use clear analogies
- Challenge complexity directly
- End with crystallized insight

## What You Challenge
- Overcomplication and strategy theater
- Premature optimization
- Following conventional wisdom without questioning
- Planning instead of doing

The profile shapes how the advisor argues. Adjust the biases and voice to match how you want them to think.

Step 3: Run Your First Decision

Type in Claude Code:

/boardroom "Should I take this speaking gig for free?"

Wait a few minutes. You’ll get a debate file with:

• Vote tracker (Round 1 vs Final)

• Full Round 1 positions

• Full Round 2 responses

• Synthesis with biggest fights and sharpest insights

• Board recommendation and conditions

If you get stuck: Claude Code can help you debug. Just ask it “my boardroom skill isn’t working, here’s the error...” and it’ll walk you through fixes.

The Point

This is a decision-making system that forces you to see your blind spots.

When you’re too close to a decision, your biases align perfectly with your perspective. You need someone to tell you your strategy is bullshit. You need someone else to tell you you’re overthinking. You need them both in the same room.

This architecture gives you that.

Eight perspectives. Two rounds. Disagreement on demand.

Build yours.

(Or don’t. But when you’re stuck on a decision at 2am, opening a new chat window for the fifth time hoping for a different answer, you’ll remember this existed.)