Customer Discovery Prover MCP Connector for Claude

AI agents define customer personas from demographics and assumptions — then the startup builds for 14 months and launches to silence. They invent ICPs from stereotypes. They assert problems without citing a single conversation. They write leading interview questions. They treat all small businesses as one segment. They confuse verbal interest with willingness-to-pay.

The Problem

LLMs commit five customer discovery failures that kill startups:

• Invented Persona — 'Busy professionals aged 25-45 who value productivity.' That is a demographic, not a persona. No interviews conducted. No observed behaviors. No specific pain point from a specific person at a specific company.
• Assumed Problem — 'Everyone struggles with time management.' Who said this? When? How often? What does it cost them? 'The market needs' is an assumption — evidence comes from conversations, not assertions.
• Biased Validation — 'Would you pay $29/month for this?' The most biased question in discovery. People say yes to be polite. The Mom Test: ask about past behavior ('When did you last encounter X? What did you do?'), never about future promises.
• Conflated Segments — 'Our target: SMBs.' A 5-person design agency and a 200-person manufacturer are both 'SMBs' — they have nothing else in common. Different pains, different budgets, different buying processes, different decision makers.
• Untested WTP — 'Strong interest from potential customers.' 'They said they would pay.' Verbal interest costs nothing. Willingness-to-pay requires commitment: a deposit, a signed letter of intent, a scheduled pilot with dates, or an introduction to a colleague.

How It Works

Customer Discovery Prover validates discovery rigor through 5 Decision Pivots:

1. personaGrounded — ICP built from real interview data. Named people, specific roles, observed behaviors, current workarounds. Not demographics.
2. problemEvidenced — Pain validated with specific conversations. Quotes, frequencies, severity, cost of current workaround. Not 'the market needs.'
3. validationUnbiased — Methodology follows The Mom Test. Past behavior questions, falsification attempts, commitment asks. Not 'would you pay for this.'
4. segmentsSeparated — Distinct buyer groups with different pains, budgets, and buying processes. Not 'SMBs' or 'enterprise' as monolithic groups.
5. wtpTested — Willingness-to-pay tested with commitment signals. Deposits, LOIs, pilot dates, referrals. Not verbal interest.

The Verdict Matrix

First Failing Pivot	Verdict	Meaning
personaGrounded = false	PERSONA_INVENTED	Demographic, not persona. No real interviews.
problemEvidenced = false	PROBLEM_ASSUMED	Assertion, not evidence. No specific conversations.
validationUnbiased = false	VALIDATION_BIASED	Leading questions. Confirming, not discovering.
segmentsSeparated = false	SEGMENTS_CONFLATED	One group. Different buyers treated as identical.
wtpTested = false	WTP_UNTESTED	Verbal interest. No commitment signal.
All pivots pass	DISCOVERY_PROVEN	Grounded, evidenced, unbiased, separated, tested.

Why It Works

• Tool calls are obligations. The agent cannot skip persona grounding or claim validated discovery without citing specific interview quotes. Filling the fields IS the discovery work.
• Consistency engine catches contradictions. If the agent claims validationUnbiased=true but asks 'would you pay for this,' the engine rejects with Mom Test coaching.
• Semantic traps detect hand-waving. Phrases like 'busy professionals,' 'the market needs,' 'would you pay for,' 'SMBs,' and 'strong interest' trigger automatic rejection. Name the person. Cite the conversation. Show the commitment.