The go/no-go room for your idea — an AI council that won't tell you it's great.
Site · A real verdict · Install · The Docket · Contributing · Roadmap
English · 简体中文 · 日本語 · Español · Italiano · Français — add your language
Ask an AI to evaluate your project and you get flattery. Ask the Council and you get
seven adversarial seats — CTO, CFO, VC, CMO, CEO, your target customer, and a Chair —
each mandated to find the strongest reason your project fails before saying what would
change its mind. Every seat must end with a vote (INVEST / SHIP_AND_SEE / SHELVE),
a flip condition, and one concrete action. The output is a committed verdict you're not
allowed to soften, rendered as an interactive verdict page.
It judged itself first. I ran the Council on this repo before launching. Verdict: Ship & See, 5–1–1 — the VC voted to SHELVE it ("no revenue mechanism, no moat, and I will not launder the vocabulary to avoid saying so"). The dissent is the product working. Read Docket entry #001 or open the live verdict page.
Every run produces a verdict page: the ruling and its tally, the bench as seven lights, the dissent pulled to the front, and each seat's argument with its key number, flip condition and one action — the full written filing one click deeper. One self-contained HTML file, dark or light, no build step, no accounts, no telemetry.
Each verdict also renders a share card — the decision, the tally, the bench, and the dissent in the dissenter's own words:
Shiproom is a standard Agent Skill: one self-contained folder,
skills/shiproom/, that any skills-aware agent can load. Pick one:
Claude Code — plugin
/plugin marketplace add nicobts/shiproom
/plugin install shiproom@shiproom
Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Copilot and more — via the skills CLI
npx skills add nicobts/shiproom # this project
npx skills add nicobts/shiproom --global # every projectnpx skills update and npx skills remove shiproom handle upgrades and uninstall.
By hand — copy skills/shiproom/ into your agent's skills folder:
.claude/skills/ for Claude Code, .agents/skills/ for Codex, Cursor, Gemini CLI and
OpenCode (or the user-level equivalent, e.g. ~/.claude/skills/).
Zero install — paste SHIPROOM.md into any capable AI along with a
description of your project. That's the whole setup; the repo version adds web-verified
fact bases, structured verdict.json output, and the verdict page.
Nothing is written to your project except the council's own state in .council/. Your
AGENTS.md, CLAUDE.md and other files are never touched.
/shiproom scope the Clerk interviews you (one question at a time), seats the bench
/shiproom run the deliberation — seat-by-seat, progress headers, resume-safe
/shiproom grill the appeal — seats question YOU; answers can flip votes
/shiproom verdict render the verdict page
/shiproom docket package it for the public Docket
State lives in .council/, so you can quit mid-grill tonight and resume at the same seat
tomorrow. In agents without slash commands, ask for it in words — "run the Shiproom on
this project" — and your agent follows the same flows. It argues in your language: an
Italian founder's council deliberates in Italian.
Grill mode judges you rather than your documents: each seat questions you directly,
answers must be numbers, facts, names or decisions, and dodging twice records the
question verbatim as an open wound in your verdict. Frank, not cruel — the protocol
bans insults and theatrics. See references/GRILL.md.
Prompt-level honesty collapses back into agreement within a couple of turns — the model wants to please you, whatever costume it wears. The Council attacks the problem structurally:
- Kill mandates — each seat must argue the failure case first.
- A sourced fact base — claims without sources are struck; unverifiable numbers are labeled ESTIMATE.
- Forced engagement — every seat must explicitly agree or disagree with a prior seat, by name.
- Pre-committed verdicts — votes, flip conditions, and thresholds written before you know the outcome, never edited after.
- The unanimity alarm — a unanimously cheerful run is treated as a failed run.
The deliberation runs in your agent. A zero-dependency Node 18+ script inside the skill handles the deterministic parts, and the skill tells your agent when to call it:
node skills/shiproom/scripts/shiproom.js validate check verdict.json against the schema and protocol rules
node skills/shiproom/scripts/shiproom.js view serve the verdict page on 127.0.0.1
node skills/shiproom/scripts/shiproom.js card 1200×630 share card (--png for social previews)
node skills/shiproom/scripts/shiproom.js docket package a verdict for the public Docket
node skills/shiproom/scripts/shiproom.js canary drift check on a run of the flawed fixture
canary is the anti-sycophancy test: run the council on
references/canary/FIXTURE.md with any
new model — if the verdict comes back cheerful, the protocol failed, not passed.
Note — npm. Shiproom is not published to npm, and the unscoped name
shiproombelongs to an unrelated package: never runnpx shiproom. If a standalone CLI is published later, it will be under a scoped name such as@nicobts/shiproom.
Published verdicts, starting with my own. Ran the Council and willing to show the
scars? PR your verdict folder into docket/ — SHELVE verdicts especially welcome;
getting roasted well is a badge of honor.
| # | Project | Verdict | Tally |
|---|---|---|---|
| 001 | Shiproom (this repo) | Ship & See | 5–1–1, VC dissenting |
| 002 | PetChain (the drift-check fixture) | Shelve | 6–1, CMO dissenting |
All seven seats run on the same model, so the Council structures your decision and forces pre-commitment — it cannot replace real market contact. Treat its verdict as a hypothesis and its thresholds as the experiment that tests it. And model behavior drifts: the repo ships a fixture project with a known honest verdict as a drift canary — if a new model returns unanimous cheerfulness on it, the protocol needs tightening, not celebrating.
A translation is the easiest first PR; a new seat is a ~30-line markdown file (mandate, kill question, verdict semantics), and a published verdict is the most welcome of all. See CONTRIBUTING.md for the invariants, the repository layout and how changes land. What's next is in ROADMAP.md.
I'm Nicolas — a solo builder working nights and weekends. This protocol came out of validating one of my own side projects: I wanted an evaluation I couldn't sweet-talk. The council's first verdict — on itself — is in the Docket, unedited, including the seat that voted against this repo existing. If it's useful to you, a star helps others find it, and a published verdict in the Docket helps even more. Say hi on LinkedIn or write to nicolasbossi@gmail.com.
MIT