Idea Validator

A Claude Code skill that pressure-tests your startup, product, or feature idea, then tells you whether to pursue it and what to test first.

It runs a panel of named sub-agents who argue both sides, finds the one assumption everything rests on, and designs an offer you could sell today — all written up in one linked report.

When you're excited about an idea, you're the worst judge of it. This skill gives you the room of skeptics and believers you'd want before you commit, each working independently so none of them can lean on your hope or on each other's conclusions.

Inspired by Garry Tan's gstack plan-ceo-review: the same adversarial, multi-agent, opinionated spirit, pointed at validating ideas instead of reviewing code.

---

What it does

Run /validate-idea <your idea> and the skill:

1. Sharpens the idea into a testable frame (problem, customer, today, proposal, why-now, wedge).
2. Picks a mode with you — Gut Check, Full Gauntlet, Kill Mode, or Investor Mode.
3. Runs the dialectic — two personas in parallel: Sloan the repeat founder builds the strongest honest case for; Dutch the burned operator builds the strongest honest case against. Same idea, opposite jobs.
4. Runs the specialist panel — more parallel personas, each adversarial in its lane:
   • Priya (market) — is the budget real, big enough, why now (web-researched)
   • the Customer — the actual buyer, in first person, "I'm busy and I already have a way to do this"
   • Marcus (competitor) — the incumbent who copies it in a quarter (web-researched)
   • Reggie (operator) — can it be built, distributed, and paid for
   • the Coroner (pre-mortem) — the autopsy: how it's dead in 24 months
   • Gabriela (investor) — in Investor Mode only
5. Distills "What Has To Be True" — turns the believer's case and the skeptic's objections into the explicit list of conditions that must all hold to win, ranked by how shaky each one is. The shakiest is the riskiest assumption.
6. Finds the cheapest test — designs the experiment that de-risks the shakiest assumption, with kill and success criteria.
7. Renders a verdict — PURSUE / PURSUE-REFRAMED / PARK / PASS, scored against a rubric where one fatal flaw caps the whole score.
8. Designs the offer — The Closer builds an irresistible offer and a cash-flow-smart price (deposit-for-credit, founding pricing, guarantees) where the paid pre-sale is the validation test. Money beats words.
9. Red-teams its own verdict — The Adversary challenges the conclusion in both directions (too kind? too harsh?), and the verdict gets revised or defended in the open.
10. Writes a linked report — a directory of files (executive-summary index → dialectic, what-has-to-be-true, panel, the test, the offer, red team) plus ready-to-use artifacts: a landing-page prompt, an interview script, and cold outreach.

You make every important call. The skill recommends hard, then leaves the decision to you.

flowchart TD
    idea([Your idea]) --> sharpen[Sharpen into a testable frame<br/>problem · customer · today · proposal · why-now · wedge]
    sharpen --> mode{Pick a mode<br/>Gut Check · Full Gauntlet · Kill Mode · Investor Mode}
    mode --> dialectic

    subgraph dialectic [The dialectic · parallel · opposite jobs on identical input]
        direction LR
        sloan[Sloan · repeat founder<br/>strongest honest case FOR]
        dutch[Dutch · burned operator<br/>strongest honest case AGAINST]
    end

    dialectic --> panel

    subgraph panel [Specialist panel · parallel · blind to each other]
        direction LR
        priya[Priya<br/>market]
        customer[the Customer<br/>the buyer]
        marcus[Marcus<br/>competitor]
        reggie[Reggie<br/>operator]
        coroner[the Coroner<br/>pre-mortem]
        gabriela[Gabriela · investor<br/>Investor Mode only]
    end

    panel --> whtbt[What Has To Be True<br/>conditions ranked by shakiness]
    whtbt --> risk[Riskiest assumption<br/>the shakiest condition]
    risk --> test[Cheapest test<br/>kill + success criteria]
    test --> verdict{{Verdict<br/>PURSUE · PURSUE-REFRAMED · PARK · PASS}}
    verdict --> closer[The Closer<br/>irresistible offer = the validation test]
    closer --> adversary[The Adversary<br/>red-teams the verdict, both directions]
    adversary --> report[(Linked report + artifacts)]

Loading

---

Install

./install.sh          # symlinks the skill into ~/.claude/skills/validate-idea

Then /validate-idea works in any session. Uninstall: rm ~/.claude/skills/validate-idea.

Reports land in idea-validations/<date>-<slug>/ in your current working directory, so you can watch the files appear live as the panel runs.

---

Usage

/validate-idea an AI chief-of-staff for solo founders that reads my email,
calendar, and Slack and proactively runs my day

or conversationally: "I'm thinking about building X. Pressure-test it before I commit."

Modes you can ask for directly:

• "just gut-check this" → Gut Check (fast)
• "tear it apart" / "I'm already in love with this" → Kill Mode
• "I'm pitching VCs" → Investor Mode
• otherwise → Full Gauntlet (default)

Swap personas on the fly: "be my toughest customer Dana" or "review it like Bezos."

---

A worked example (real results)

skills/validate-idea/examples/2026-06-27-ai-chief-of-staff-solo-founders/ is a real run of the full gauntlet — the panel produced it live with real web research. Start with index.md (the 60-second executive summary) and follow its links. The verdict came out PURSUE-REFRAMED at low confidence (effectively a PASS as framed), with a $500 concierge pre-sale as the validation test. Fatal flaws and all — that's the real output.

---

Repo layout

idea-validator/
├── README.md
├── install.sh
└── skills/validate-idea/
    ├── SKILL.md                 # the skill — read as instructions
    ├── references/
    │   ├── panel.md             # the named personas + return contracts
    │   ├── rubric.md            # scoring + capping rules + verdicts
    │   ├── offer-playbook.md    # business/offer knowledge (the value equation, cash-flow plays)
    │   └── report-format.md     # the linked multi-file output spec
    └── examples/
        └── 2026-06-27-ai-chief-of-staff-solo-founders/   # a real produced run

---

Design notes

• The disagreement is built in. Sloan and Dutch get opposite jobs on the same input. The five lenses can't see each other's work, and the Adversary can't see the panel. That independence is the whole point, so each persona runs as its own sub-agent with its own context.
• "What has to be true" turns argument into action. The believers and skeptics each imply a set of conditions. List them, rank them by how shaky they are, and the shakiest one becomes the bet you test first.
• One fatal flaw caps the score. Average the scores and a weak idea slips through with a passing grade. So a single killer finding — no market, no moat — overrides an otherwise high total.
• Money is the real test. A paid deposit from the exact buyer is the strongest signal you can get. The offer is built to be that experiment, with a number that tells you when to walk away.
• You're in charge. The skill recommends hard, then hands every real decision back to you.