boardroom
datysho/boardroom
The AI board that keeps score on itself — a virtual board of directors for business decisions, with Brier-scored pre-registered predictions
Instalação
npx convez-ai datysho/boardroom1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
BOARDROOM
The AI board that keeps score on itself.
Your investors have a board. You don't. Every serious call you make — which project gets the next quarter, which idea to kill, which hire to trust — happens alone, with no dissent on record and nobody keeping score. BOARDROOM is a Claude Code plugin that convenes eight directors with conflicting mandates, makes them argue your decision, writes the verdict down with predictions attached, and then scores every prediction — including yours.
What it is
A decision protocol with a journal and a scoreboard — not a chat. One question moves through:
- Eight directors with conflicting mandates. The Visionary may not discuss feasibility; the Skeptic may not open with praise; the Customer must be a busy, skeptical buyer. Disagreement is designed in, not prompted for.
- An independent first round. Directors write their positions blind — they never see each other before peer review, and scoreboard data never reaches them, so there is no anchoring and no favorite to follow.
- Adversarial peer review. Positions are anonymized (labels A–H), each director's own memo is their prior, and changing position requires citing a specific piece of evidence. If four or more seats drift toward one option, the memo carries a HERDING FLAG and post-review votes count for less.
- A decision memo with pre-registered predictions. Every verdict ships with the vote, the preserved dissent, tripwires, and dated falsifiable predictions (
p = NN%) — written before reality answers. - A scoring ritual that scores everyone.
/boardroom:scoreBrier-scores every due prediction against the seat's own stated criteria — and when you override the board, your OWNER ACTION is scored by exactly the same rules. The board keeps score on itself, and on you.
The Boardroom Loop
BRIEF ──> DISSENT ──> DECIDE ──> TRACK ──> SCORE
▲ │
└────── every score sharpens the next ─────┘
brief
- BRIEF — idea intake and the secretary's sweep produce a falsifiable Decision Brief.
- DISSENT — research, eight independent positions, adversarial peer review.
- DECIDE — the memo: vote, preserved dissent, tripwires, pre-registered predictions, your OWNER ACTION.
- TRACK — between sessions: tripwires checked against facts, overdue reviews surfaced, the journal kept.
- SCORE —
/boardroom:score: Brier scoring for every seat, and for you.
Full mechanics, gates, and quorum rules: METHOD.md.
It argues from your record
As your journal grows, the board starts citing it back at you: "your last four overrides of the Skeptic went 1-for-4", "you call this project the focus, but you logged 2 hours on it last month", "this is the third decision matching your blind spot #2", "of the board's own KILL calls, one was a miss — here is which". Every such line cites a file, a date and a count that exist on your disk — and when there is no record yet, the board says nothing. A tool that prints lines like these on day one is making them up.
The seats
| Seat | The job title it replaces | Model |
|---|---|---|
| visionary | The venture partner asking "how big can this possibly get?" | opus |
| skeptic | The independent director who red-teams every plan | opus |
| cfo | Your fractional CFO: unit economics, cash, opportunity cost | sonnet |
| operator | The COO: the 90-day plan, who exactly does the work, what breaks at 10x | sonnet |
| growth | The CMO who names the first 100 customers, channel before product | sonnet |
| counsel | The startup lawyer on retainer: regulation, claims, IP, jurisdiction | sonnet |
| historian | The analyst who knows the base rates: what happened to the last hundred | sonnet |
| customer | The busy, skeptical buyer from your ICP: would they actually pay? | haiku |
| scout (staff) | The research associate: facts with sources, no opinions | haiku |
That is not a chat. That is a board.
Quick start
1. Install the plugin.
claude --plugin-dir /path/to/boardroom
Commands appear as /boardroom:...; the secretary attaches automatically and quietly flags overdue reviews at session start.
2. Run /boardroom:setup.
A short interview — one question at a time — and the Chairman writes the board's memory for you: who you are, your goals, one file per project, your owner profile. You never touch raw config. Upgrading from an older board? Setup migrates your old files into the new context tree and archives the originals.
Want to see a finished board in 10 minutes? In setup, choose the demo company: three competing projects and two past decisions whose predictions come due immediately — run /boardroom:score and watch real Brier scoring fill the scoreboard. Every demo file is clearly marked as a demo.
3. Ask your first question.
/boardroom:ask "Should we raise prices 20% for new customers?"
Three directors, one round, a verdict in minutes. When a one-way-door decision lands on your desk, graduate to /boardroom:decide.
Receipts
We don't have testimonials. We have a journal.
- We ran the board on this product. Before launch, BOARDROOM evaluated the decision to launch BOARDROOM — verdict and dissent included, predictions pre-registered.
- PUBLIC-TRACK-RECORD.md — every public prediction this project makes, pre-registered with probabilities before the event, scored quarterly. Git history is the proof: each prediction is committed before it resolves.
Commands
| Command | What it does |
|---|---|
/boardroom:decide "question" | Full board on one decision: research, independent positions, peer review, a decision memo with scored predictions. Two-way doors get a short 5-seat protocol. |
/boardroom:pitch "idea" | Evaluate a new business end-to-end: PURSUE / PILOT / PARK / KILL, gated by ambition (ceiling class) and discipline (survival + your specific asset). |
/boardroom:focus | Project table (no arguments) or the full quarterly focus session with a Hamming round: "why aren't you working on the most important project?" |
/boardroom:risks "plan" | Six directors assume the plan already failed and write the obituary; the Chairman clusters it into top risks with early-warning signs. |
/boardroom:ask "question" | Skeptic + CFO + Growth, one round, verdict in minutes. Logged; a question that returns a third time gets escalated to decide. |
/boardroom:score | The scoring ritual: Brier-score every due prediction by its own pre-registered criteria — every seat's and yours — update the scoreboard and your owner profile. |
/boardroom:alarms | Audit every active decision's tripwires against the facts; fired ones reopen the decision. |
/boardroom:idea | Six forcing questions, one at a time, that turn a raw thought into a board-ready brief with a falsifiable success criterion. |
/boardroom:rival "rival" | Play the incumbent: two rounds of move / counter-move, ending in a response matrix and pre-registered predictions about competitor behavior. |
/boardroom:hire "role/person" | Evaluate a key hire: role scorecard, base rates, halo-effect red team, true cost — with 90-day scored predictions. |
/boardroom:history "topic" | Digest everything the board ever decided about a topic into a prior-history brief: verdicts, prediction outcomes, lessons. |
/boardroom:library | Digest books and materials you drop into the library; directors quote seat-tagged digests, never originals. Digests stay in your private board folder and never enter any public artifact. |
/boardroom:status | One map of the whole board: due reviews, fired tripwires, the focus vs capacity — ending in exactly one recommended next command. |
/boardroom:setup | Guided onboarding interview; writes the board's memory files for you. Optional demo company; migrates older boards. |
Renamed in v4
Same board, human names. A clean break — the old names are gone, no aliases.
| Old | New |
|---|---|
| convene | decide |
| venture | pitch |
| portfolio | focus |
| premortem | risks |
| quick | ask |
| review | score |
| check-tripwires | alarms |
| intake | idea |
| wargame | rival |
| keyhire | hire |
| recall | history |
setup and status keep their names. The vocabulary moved with the commands: what older texts called a bet is now a project, and the main bet is now the focus. The verdicts — PURSUE NOW / PILOT / PARK / KILL — and the kill criteria stay blunt and untranslated. Signed memos in the track record keep their original wording as historical record, with dated addenda instead of rewrites.
What it costs
No API key and no separate bill — every director runs inside your Claude Code subscription. Honest numbers:
| Command | Director runs | Typical time (Max plan, estimated*) |
|---|---|---|
| decide — full board (one-way door) | 19 (4 opus, 10 sonnet, 5 haiku) | 15–25 min |
| decide — short protocol (two-way door) | 5–8 | ~10 min |
| pitch | 20–22 (4 opus, 10–12 sonnet, 6 haiku) | 20–30 min |
| focus — quarterly session | 20 (5 opus, 11 sonnet, 4 haiku) | 20–30 min |
| focus — table mode | 0 | ~1 min |
| rival — 2 rounds | 7 (2 opus, 4 sonnet, 1 haiku) | 10–15 min |
| risks | 6 | 5–10 min |
| hire | 4 (1 opus, 3 sonnet) | 5–10 min |
| ask | 3 | minutes |
| score, alarms, idea, history, library, status, setup | 0 | the Chairman works alone |
*Times are estimates pending measured public sessions; run counts are exact (derived from the seats table).
Every multi-director command shows this plan up front and asks once: full / short / quick. On a Max plan a full session is routine load; on a Pro plan the board itself recommends the short protocol or /boardroom:ask.
State & privacy
Everything the board knows lives in one folder: ~/boardroom (or $BOARDROOM_HOME if you set it).
~/boardroom/
context/
owner.md who you are: facts, assets, hard constraints
values.md values and anti-goals — read by the Chairman only
goals.md goals by horizon (1y / 3y / 10y): a number and a date each
now.md capacity, the named focus, what changed lately
jurisdiction.md legal context — read by the counsel seat only
projects/ one project = one file, kill criteria from birth
library/ books and materials the board actually reads
owner-profile.md revealed risk appetite, overrides, blind spots
scoreboard.md calibration for every seat — and for you
decisions/ decision memos + INDEX.md
log.md one line per quick check
- Plain files you can open, edit, and delete in any editor.
- Projects are files. Each project is one page: sized in engineer-weeks and dollars, with kill criteria from birth and a review date. The named focus lives in
now.md. - The library is books the board actually reads. Drop a PDF, EPUB, DOCX, or text file into
library/originals/; on your explicit command the board digests it — claims and numbers, tagged by seat — and directors quote the digest, never the original. Digests stay private: they never enter this repository or any public artifact; like every board file, their content reaches Anthropic for inference when directors quote them. - No API keys — it runs entirely inside your subscription.
- Full security and privacy posture, including the session-start hook disclosure: SECURITY.md.
- No third-party endpoints, no analytics, no phone-home. Your files go only where your Claude conversations already go — to Anthropic for inference, under the subscription terms you already accepted. Nothing in this plugin sends data anywhere else.
- MIT licensed.
What we refused to build
- No swarms, queens, or consensus topologies — eight mandates in conflict beat a hundred agents in agreement.
- No vector database — the journal is plain text and plain-text search finds the precedents; boring on purpose.
- No invented benchmarks — every claim in this README is verifiable inside this repo in minutes.
- No confidence theater — until a seat has 5 scored predictions, every memo labels its numbers as opinions, not odds.
- No backtests on famous startups — the model already knows who won; only forward predictions count.
Why each refusal, in full: METHOD.md.
FAQ
Is this financial or legal advice? No. The board is a disciplined way to argue with yourself and keep score; the counsel seat flags questions to bring to a real lawyer, it does not replace one. The decisions — and their consequences — are yours.
Can I use other models?
Each seat declares its model at the top of its file in agents/ — change it there, or set CLAUDE_CODE_SUBAGENT_MODEL to force one model for every seat.
Can my team share one board?
Yes — git init ~/boardroom and push it to a private repository. Co-founders pull the same journal, scoreboard, and context tree. No server, no accounts.