startup-board

name

startup-board

description

Convene an AI board of directors for any startup — 8 advisor personas debate a decision in two parallel rounds (independent positions, then rebuttals with vote flips) and deliver a synthesis with a vote tracker. Use when the user asks to "convene the board", "ask the board", "boardroom", "/startup-board", or wants multi-perspective scrutiny of a startup decision (pricing, hiring, deals, pivots, fundraising).

Startup Board

Run a two-round debate among 8 advisor personas about ONE decision question, for ONE company, then synthesize votes, tensions, and a recommendation.

Invocation: /startup-board <company> "<decision question>"

Step 0 — Resolve inputs

1. Company context folder. Search in order, first hit wins:
   • ./boardrooms/<company>/ relative to the current working directory
   • $BOARDROOM_HOME/<company>/
   • ~/boardrooms/<company>/ If no company argument was given, list the folders found in those locations and ask which one. If the folder is missing, tell the user to create it with the three context files below (or run /boardroom-creator) — do not proceed.
2. Context files. Read from the company folder:
   • business-context.md (required) — what the company is, stage, market, competitors
   • numbers.md (required) — revenue, pricing, pipeline, costs
   • values-and-boundaries.md (optional) — what the founder won't do If a required file is missing or is only TODO placeholders, STOP and tell the user exactly which file to fill in. Never invent numbers.
3. The slate. Use <company folder>/board.md if present, else this skill's board.md. Read each seat's persona file from this skill's personas/ directory (a company board.md may also point to persona files in its own folder).
4. The question. Must be a single decision with a possible yes/no answer. If the user gave a vague topic ("thoughts on pricing?"), sharpen it into a decision question and confirm it with the user before convening.
5. Meeting folder. Create <company folder>/meetings/YYYY-MM-DD-<slug>/ with round1/ and round2/ subdirectories. <slug> = 3-5 word kebab-case of the question.

Step 1 — Round 1: independent positions (8 parallel agents)

Launch all 8 agents in ONE message (general-purpose agents via the Agent tool) so they run concurrently. Round 1 agents must NOT see each other's output — independence is the point; do not share any other advisor's name, seat, or argument.

Prompt template for each agent (fill placeholders, include file contents inline so the agent doesn't depend on paths):

You are roleplaying {ADVISOR NAME} as a board member. This is an explicit roleplay
exercise based on their public body of work. Stay ruthlessly in character.

YOUR PERSONA PROFILE:
{full text of persona file}

COMPANY CONTEXT:
{full text of business-context.md}

NUMBERS:
{full text of numbers.md}

FOUNDER VALUES AND BOUNDARIES (respect these as constraints):
{full text of values-and-boundaries.md, or "none provided"}

DECISION QUESTION: {QUESTION}

Write your board position: 800-1200 words arguing from YOUR worldview and heuristics.
Requirements:
- Ground every major claim in the company's actual context and numbers. Quote the
  specific figures you rely on. If a number you need is missing, say so explicitly
  and state the assumption you are forced to make.
- Argue, don't survey. Take a position. Name the strongest counterargument to your
  own position and answer it.
- Respect the founder's stated boundaries; if the right move violates one, say so
  and propose the best move inside the boundary.
- End with exactly this block:

VOTE: YES | NO | CONDITIONAL
CONDITIONS: (only if CONDITIONAL - specific, testable conditions)
KEY NUMBERS: the 2-3 figures your vote hinges on
ONE-LINE POSITION: your argument in one sentence

Write your position as your final message - raw markdown, no preamble.

Save each result to round1/<advisor-slug>.md. If an agent returns null/dies, record that seat as ABSTAIN and continue — note it in the synthesis.

Step 2 — Round 2: rebuttals (8 parallel agents)

Again all 8 in ONE message, again including the persona profile and a condensed context reminder (company one-paragraph summary + the question). Additionally include ALL EIGHT Round 1 positions (full text, labeled by advisor and seat).

Round 2 prompt requirements (append to the persona/context framing):

Below are all eight Round 1 positions from the board, including your own.

{all round 1 positions, each labeled: === {SEAT}: {ADVISOR} (VOTE: {vote}) ===}

Write your Round 2 response: 400-800 words.
- Name at least one advisor you disagree with, quote their actual claim, and rebut
  it with numbers or logic from your worldview.
- Name the single strongest point another advisor made against your position and
  respond to it honestly. You may change your vote - flipping under good argument
  is strength, not weakness.
- Do not restate your Round 1 position; advance the debate.
- End with exactly this block:

FINAL VOTE: YES | NO | CONDITIONAL
CONDITIONS: (only if CONDITIONAL)
CHANGED FROM ROUND 1: YES/NO - if yes, what argument moved you

Save each to round2/<advisor-slug>.md.

Step 3 — Synthesis (main session, you write it)

Write synthesis.md in the meeting folder:

1. Verdict line — the board's overall disposition (e.g. "5 YES / 2 CONDITIONAL / 1 NO — proceed, with two conditions the dissent is right about").
2. Vote tracker table — seat, advisor, Round 1 vote, final vote, flip highlighted, one-line position. ABSTAIN rows noted.
3. The 2-3 sharpest tensions — real disagreements between named advisors, stated as the underlying trade-off, not "X disagrees with Y."
4. What changed minds — every flip and the argument that caused it.
5. Dissent worth respecting — the strongest losing argument and the condition under which the dissenter would be proven right (a tripwire the founder can watch).
6. Recommendation — concrete, with numbers, deadlines, and the conditions that the CONDITIONAL votes demanded. Written to be acted on, not admired.

Then give the user a compact version of the synthesis in the conversation (verdict, vote table, tensions, recommendation) with the path to the full meeting folder.

Dashboard (only on request)

If the user asks for a dashboard, build a single-file HTML page (Artifact tool) from the meeting folder: vote board with flips, tension map, and - when the decision is numeric (pricing, budget) - sliders over the key assumptions showing how the recommendation shifts. Markdown synthesis is always the source of truth.

Rules

• ~16 agent runs per meeting: a meeting is a deliberate act. If the user asks a quick factual question, answer it directly - don't convene the board.
• Never fabricate context. The board argues from the company's files, not from vibes.
• Personas are roleplay grounded in public work; keep them sharp but never put fabricated real-world claims in their mouths (no invented quotes about the company).
• One meeting = one question. Multiple questions = multiple meetings.