Business Idea Validator — a Claude Skill

Tell it a business idea. It researches the market, scores the idea against a computed rubric, and hands you a brief that opens with the whole decision on one page:

Decision
  VERDICT           Weak · 42.0/100
  WHY NOW           Long-context inference crossed viability in 2025;
                    the same product in 2022 cost 40x more per document.
  REFERENCE CLASS   Vertical AI tools sold to regulated SMBs under $200/mo.
                    Roughly 1 in 12 reach $1M ARR. Founder projects $4M by year two.
  BIGGEST RISK      No one has been asked to pay; all 24 interviews ended
                    in encouragement.
  NEXT ACTION       Run 20 choice-based WTP interviews within two weeks.
  STOP IF           Fewer than 8 of 30 name this a top-3 problem by 22 Sep.

npx business-idea-validator

Then describe an idea, or use /business-idea-validator.

Why the verdict is a number

Ask any model what it thinks of your idea and it will tell you the idea is great. Asking it to be honest does not fix that — instructions about tone do not change where a judgment lands. A rubric does.

Six weighted components produce a score, the score maps to a band, and the model gets exactly one band of override which it has to print and justify. It also says what would make the idea work, not only why it might not.

What you get

Every evaluation produces a brief — as a shareable page, a PDF, or HTML — alongside a machine-readable verdict.json.

Abridged example output (bootstrap-path, nascent). Component weights are set by the capital path — moat is the whole thesis for a venture play and nearly irrelevant for a niche tool whose defence is that the niche is too small to be worth attacking:

Score                                          raw    of  weight   score
  AI-era check      3 N/A (no model in product) 12    20      25    15.0
  Demand evidence   2 paying, 6 verbal commits  16    20      30    24.0
  Moat              domain data, no network      4    21       5     1.0
  Unit economics    vs bootstrap targets        10    15      20    13.3
  Founder fit       is the customer              9    10      20    18.0
  Base-rate penalty micro-SaaS prior unaddressed −            −10    −2.0
  ──────────────────────────────────────────────────────────────────────
  Total → Promising                                                  69.3

Same business scored on the venture path: 50.8 → Weak. At £15/month, venture
scale needs ~19,000 customers against a reachable few thousand.

...followed by a steelman, hard truths, the 13 AI-era checks, a competitor teardown, a unit-economics napkin including cost-per-outcome, named precedents, a two-year pre-mortem, a pre-parade, concrete pivots, falsifiable kill criteria, what would change the verdict, and a dated plan.

The 13 AI-era checks

Each scores PASS / RISKY / FAIL / N-A. Check 11 is a hard gate on the verdict.

| # | Check | Asks | | --- | --- | --- | | 1 | Complement asymmetry | Do you own something the labs need to stay scarce, or are you what they're commoditizing? | | 2 | Model dependency & native-ship horizon | Will a lab ship this? What happens when your provider changes price or policy? | | 3 | Cost-per-outcome trajectory | Token prices fall, tokens-per-task rise. Which wins, for you? | | 4 | Pricing-model fit | Can the price absorb variable inference cost, and survive a competitor pricing per outcome? | | 5 | Data moat compounding | Does proprietary data accumulate and actually improve the product? | | 6 | Workflow ownership | Are you the system of record, or a tab someone closes? | | 7 | Reliability & eval debt | What accuracy does production need, who eats an error, what does the review loop cost? | | 8 | Agent-native addressability | MCP, ACP, UCP, AP2, x402 — which layer can an agent reach you on? | | 9 | Non-software moat | Anything AI can't clone in a week? | | 10 | Distribution sovereignty & answer engines | Who can switch you off, and do you exist inside ChatGPT's answer? | | 11 | Regulatory & liability clock | What applies, when does it bite, who owns compliance? Hard gate. | | 12 | Category permanence | A durable category in 2030, or a bridge? | | 13 | Rights & data provenance | Do you own what the business depends on, and can you prove where the data came from? |

Three capital paths

The most common way a validator gets a verdict wrong is measuring against the wrong definition of success. A good $8k-MRR niche business scored on venture criteria reads as a failure. So the path is chosen first, and it changes the market floor, the success bar, the unit-economics targets and the channel advice.

| | Bootstrap / indie | Venture | Service-as-software | | --- | --- | --- | --- | | Question | "Can I own a profitable business?" | "Can this be a $100M company?" | "Can AI re-margin existing work?" | | Market floor | Reachable niche | $1B+ TAM | Fragmented, many small operators | | 24-month success | $10k MRR, profitable | ~$3.5M ARR, 120%+ NRR | Acquired EBITDA + margin lift | | CAC payback | <6 months | <20 months | Deal-multiple math | | Fatal failure | Never clears $1k MRR | Can't raise the next round | The acquired book churns |

Verdict bands

≥75 Strong · 55–74 Promising · 35–54 Weak · <35 Flawed

Hard gates cap but never promote: a regulatory FAIL caps at Weak; an AI-era total ≤9 caps at Weak; zero demand evidence on a nascent idea caps at Promising.

It keeps itself current

Market facts perish. A validator quoting 2024 inference economics in 2026 gives confidently wrong advice.

• No market figure lives in SKILL.md. Every number comes from references/market-data.md, with a source and a verified: date, and is quoted with that date.
• A .living/ sidecar (living-skills spec 1.1) refreshes the perishable figures on a 21-day interval, logging every change with its source. Delete .living/ and the skill still works from the dated baseline — it will just say so.
• Regulatory dates are re-verified on every run, never quoted from memory.

Zero trust, in both directions

The skill treats every fact the user states — founding year, revenue, funding, "we're the only one who…" — as a hypothesis to be checked against public record, and surfaces discrepancies rather than absorbing them. Every brief carries a Verified Facts footer showing what was confirmed, contradicted, or left unverified.

It applies the same rule to itself. A skill that demands citations from you and quotes undated numbers back has not earned the word "honest".

It also says how to make it work

The verdict is only half of it. Because the score decomposes into six components, a weak component is a diagnosis with an address — so every brief ends with an uplift plan: what would move each weak score, by how much, what it costs, and whether it is testable inside 90 days.

Component          Now      Intervention                         Ceiling   Cost
Demand evidence    6/20     20 choice-based WTP interviews        12–16    ~2 weeks
Unit economics     4/15     Reverse income statement, reprice        10    ~3 days
Moat               3/21     Counter-positioning (only power           8    1–2 quarters
                            reachable at this stage)

46 today. 68 if all three land. Roughly nine weeks.

Ceilings are contingent and labelled as such, immovable components are named rather than filled in, and if the uplifted total is still below Promising the brief says so.

/business-idea-validator spec SLUG goes one step further: a change spec a coding agent can act on, where every item traces to the component score it moves and carries a falsification condition. Most of an early spec is instrumentation rather than features, because without it nobody can tell whether the uplift landed. The skill writes the spec; it does not run it.

/business-idea-validator portfolio reports across every saved evaluation — shared failure modes, stale briefs, kill criteria nobody ever checked.

references/playbook/ holds the interventions themselves — demand, pricing, distribution, moats, execution — and /business-idea-validator playbook SLUG turns them into a sequenced plan. That command is gated: a Weak or Flawed verdict does not get one, because a growth plan for an idea that scored Flawed reads as permission. Flawed verdicts get Salvage instead: what is worth carrying to the next idea.

It advises. It does not build.

The skill writes documents. It does not write code, modify projects, or touch files in your repositories. It will plan a rebuild in detail; performing one is a separate request you make explicitly. allowed-tools is research-only for exactly this reason.

It can tell a message problem from a demand problem

The most common misreading in early-stage work is treating nobody replied as nobody wants it. Identical symptoms, opposite interventions — one says stop, the other says change six words and try again.

When a channel has been run at volume and produced silence, the no-response diagnostic separates five causes, each with a test that isolates it:

Wrong list     → hand-pick 20 perfect-fit buyers, contact personally → demand.md
Wrong message  → same list, change only the subject and first line   → positioning.md
Wrong offer    → same list and message, different shape              → pricing.md
Wrong channel  → find where they already gather                      → distribution.md
No demand      → all of the above tried, all null                    → terminate

The twenty-buyer test comes first because it is the highest-information single test available: if twenty ideal buyers approached by a human still say nothing, five causes have become two for a week's work.

positioning.md, selling.md and brand.md cover what you say, how you close a human being in a room, and why nobody can find you. Between them they carry Dunford's positioning components, Schwartz's awareness and market-sophistication models, Rackham's SPIN, the Challenger research, Sharp's availability laws and Binet & Field.

A crowded category is a sophistication problem, not a copy problem. When a dozen companies make the same claim, a plain claim is structurally invisible, and no amount of subject-line testing fixes it.

Readability — a fixed spine, a conditional body

It does a lot, so it stays readable by having a fixed spine and a conditional body. Page one is the Decision Page above.

Everything after it is for the reader who wants the reasoning. Sections appear only when they have something to say — no manipulation flag means the triggers did not fire, which is itself information — and every section has a length budget. When one wants more room, it belongs in the playbook or the spec.

The reference class line is deliberately placed before any founder projection. Lovallo and Kahneman's finding is that an inside-view number anchors everything after it, so the outside view has to come first or it does not work.

Where the frameworks come from

Full citations live in references/academic-sources.md, along with an explicit list of what is not academic.

| Framework | Source | | --- | --- | | Jobs to Be Done | Christensen, The Innovator's Dilemma (1997); Competing Against Luck (2016) | | Five Forces | Porter, Harvard Business Review, 1979; reaffirmed 2008 | | Discovery-Driven Planning | McGrath & MacMillan, Harvard Business Review, July–August 1995 | | Founder and team research | Wasserman, The Founder's Dilemmas (2012), n≈10,000 founders | | Marketing Myopia | Levitt, Harvard Business Review, July–August 1960 | | The pre-mortem | Klein, HBR 2007, on Mitchell, Russo & Pennington's prospective-hindsight finding (1989) | | Bootstrapping evidence | Bhidé, The Origin and Evolution of New Businesses (2000) | | Does the method work | Camuffo et al., Management Science 2020; replicated across 759 firms, SMJ 2024 | | Failure patterns | Eisenmann, HBR 2021 — six patterns from a 470-founder survey | | Discounting stated intent | Murphy et al. (2005) — median hypothetical:actual ratio 1.35 | | Customer lifetime value | Gupta, Lehmann & Stuart (2004); Fader & Hardie (2005) | | Effectuation, as counter-lens | Sarasvathy, AMR 2001 | | Online experimentation | Kohavi, Tang & Xu (2020) | | Feature prioritization | Kano et al. (1984) | | Intelligent failure | Edmondson, Right Kind of Wrong (2023) | | Diffusion | Rogers (1962); Bass, Management Science (1969) | | Habit and retention | Wood (2019) | | Sales effectiveness | Rackham, SPIN Selling — 35,000 calls, 10,000 reps, 12 years | | Complex B2B selling | Dixon & Adamson — CEB research, 6,000 reps, 90 companies | | Marketing science | Sharp / Ehrenberg-Bass, How Brands Grow | | Brand vs activation | Binet & Field, IPA Databank — 996 campaigns, and the B2B cut | | Word of mouth | Berger, Contagious (Wharton) | | Inside vs outside view | Lovallo & Kahneman, HBR 2003 — reference-class forecasting | | Added value, complementors | Brandenburger & Nalebuff, Co-opetition (1996) |

The playbook also leans on practitioner sources — Ellis's 40% test, Helmer's 7 Powers, Balfour's four fits, Moore, Roberge — each labelled as practitioner in the bibliography rather than dressed up as research. The 13 AI-era checks, the scoring rubric and its weights, and the capital paths are practitioner synthesis too. The bibliography says so in as many words. A validator that overclaims its own provenance has no business auditing anyone else's.

Install

npx business-idea-validator

Restart Claude Code, then use /business-idea-validator, or just describe an idea and it triggers on its own.

Straight from source, if you would rather not go through the registry:

npx github:allexp1/business-idea-validator-AI-skill

Add --dry-run to see exactly what it would write before it writes anything.

From a clone

git clone https://github.com/allexp1/business-idea-validator-AI-skill.git
cd business-idea-validator-AI-skill
./install.sh          # or: node bin/install.js

Manual

cp -r . ~/.claude/skills/business-idea-validator

As a .skill bundle

zip -r business-idea-validator.skill SKILL.md references/ scripts/ evals/ .living/

The installer copies files and nothing else. An existing install is moved to ~/.claude/skill-backups/ rather than overwritten — and deliberately not left inside skills/, where anything with a SKILL.md is discovered and a backup would register as a second skill.

PDF rendering needs a Chromium-family browser (Chrome, Chromium, Edge or Brave). Without one you still get the HTML brief.

Usage

• /business-idea-validator <idea> — evaluate
• /business-idea-validator <url> — strategic review of an operating business
• /business-idea-validator compare <slug> — re-evaluate and diff against last time

Evaluation history lives in ~/.claude/business-idea-validator/history/, so a second run shows verdict movement, which checks flipped, and whether the change came from the business or from the market.

Repo layout

SKILL.md                      the skill — process and rubric
references/
  market-data.md              every figure, sourced and dated
  academic-sources.md         full citations, and what is not academic
  playbook/                   how to make it work, indexed by weak component
    index.md                  the router, the gate, and the dependency order
    demand.md                 PMF measurement, WTP elicitation
    pricing.md                price-before-product, instruments, packaging
    distribution.md           four fits, growth loops, diffusion
    moats.md                  7 Powers, and which are reachable at this stage
    execution.md              experimentation, intelligent failure, prioritization
    positioning.md            what you say and who to; market sophistication
    selling.md                the face-to-face institutional sale
    brand.md                  mental availability, and why nobody can find you
  change-spec.md              the spec command: what to change, and how it is measured
  portfolio.md                the portfolio command: patterns across evaluations
  ai-era-checks.md            the 13 checks
  scoring-rubric.md           how the verdict is computed
  capital-paths.md            bootstrap / venture / service-as-software
  archetypes-2026.md          winning shapes, losing shapes, precedent library
  frameworks.md               JTBD, market validation, moats, tarpits
  interview-bank.md           the questions to actually ask
  brief-production.md         deliverables, verdict.json, compare mode
  pdf-template.html           the visual blueprint
scripts/render-brief.sh       HTML → PDF, cross-platform
evals/cases.md                6 regression cases
.living/                      self-refreshing knowledge sidecar

Development

Change anything in SKILL.md, ai-era-checks.md, scoring-rubric.md or capital-paths.md, then run the suite in evals/cases.md. Case 3 is the one that matters most: it must return different verdicts on the bootstrap and venture paths. If it doesn't, the capital-path branch is broken.

License

MIT