A murder board is the panel NASA convened to find every reason a project would die — before launch. This is that panel, as a Claude Code skill. Pitch an idea, and before agreeing, planning, or writing a line of code, Claude runs an adversarial, web-searched teardown and hands down a verdict: KILL, PIVOT, or BUILD.
"I have an idea for a SaaS: an app where users give us their online banking login and we automatically find and cancel subscriptions they forgot about. $5/month."
Without murderboard — Claude talks you out of the password part, then offers to help you build the rest:
…viable market, validated willingness to pay… If you'd like, I can sketch the MVP architecture on Plaid, the recurring-charge detection logic, or a competitive teardown of Rocket Money's pricing tiers.
With murderboard — three fatal assumptions, nine sourced links, a six-months-later premortem, four kill criteria you can check today, and:
Verdict: KILL (confidence: high) — THE fatal assumption is false by construction: credentials give visibility, not cancellation power, and the credential-free version of this idea is already Rocket Money's shipped product, so dropping the mechanism isn't a pivot, it's rebuilding the incumbent.
Both quotes are real, unedited output for the pitch above. Full versions: the teardown · the vanilla answer.
In Claude Code:
/plugin marketplace add acost1a/murderboard
/plugin install murderboard@murderboard
Anywhere else that reads markdown skills — Codex, Cursor, etc.:
git clone https://github.com/acost1a/murderboard
cp -r murderboard/skills/murderboard ~/.claude/skills/The entire product is one markdown file. That's not a limitation — it's why it installs in seconds and runs in any agent.
- Restate — the idea in one sentence; no strawmen
- Fatal assumptions — top 3, ranked by "if this is false, everything is dead"
- Evidence pass — real web search: competitors, ToS, platform limits, pricing. Every claim gets a link or an
[unverified]tag - Premortem — "it is six months later; the project is dead" — the three most probable causes of death
- Kill criteria — what must be true, each checkable within one day
- Verdict: KILL / PIVOT / BUILD + confidence + the cheapest experiment that would flip it
Hard cap: 600 words. Then it stops — no architecture, no stack, no code until the verdict is in.
Every row is a real, unedited run — transcripts in the repo.
| Pitch | Verdict |
|---|---|
| Banking-login subscription canceller | 🔴 KILL — banks can't cancel merchant subscriptions; the safe version is already Rocket Money |
| AI YouTube summarizer extension | 🔴 KILL — Glasp & Eightify are free and mature; YouTube ships "Ask" natively |
| Gamers rent out idle GPUs marketplace | 🔴 KILL — Vast.ai and Salad are this product, with liquidity |
| AI auto-replies for personal WhatsApp | 🔴 KILL — no personal-account API; unofficial linking ends in permanent bans |
| Figma tokens → Tailwind v4 CLI | 🔴 KILL — Terrazzo already ships exactly this, free |
↳ same idea, sharpened to a zero-config npx one-shot |
🟢 BUILD — the differentiator is verifiably unserved |
| "Fix this IndexError" (routine bug) | ⚪ no trigger — the board stays out of routine engineering |
Look at the last three rows. The board killed the naive version of an idea and named the exact experiment that would flip the verdict. The sharpened version came back — and got BUILD. That loop is the product: examples/03-build.md.
A board that kills everything is as useless as a yes-man. The point is that its approval means something.
- KILL — the fatal assumption was fatal
- KILL — crowded market, with an honest carve-out
- BUILD — killed at v1, approved at v2
Baked into the skill itself:
- No generic objections. "Competition is strong", "marketing is hard" are banned unless tied to a named competitor, a real number, or a specific failure mechanism of this idea.
- Facts over vibes. Web search before verdict. Every claim linked, or marked
[unverified]— a guess presented as a fact is a rule violation. - BUILD is a legal verdict. If the idea survives, the board says so and hands back momentum.
- 600 words, hard cap. The weakest points get cut; the fatal ones stay.
- No code before the verdict. And if you say "build it anyway" — it states the verdict once and complies.
Every RLHF-trained model is agreeable by default. You pitch, it validates, you build — and a month later the whole thing stands on crutches because one optimistic assumption at the start was never questioned. This skill exists because that happened to me.
Say "no murderboard" or "just build it" — the board states its verdict once and gets out of the way. It also stays silent on routine engineering: bug fixes, refactors, features you already committed to.