---

name: founder-validation description: > Use when the user asks to validate a business idea, design experiments, draft interview scripts, capture evidence from real-world tests, or review assumption status. Phase 2 subskill of the founder family. Validation SUPPORT — designs experiments, drafts interview scripts (Mom Test), captures evidence, reads analytics via browser MCP (Envelope D). You run the experiments and report back. Modes: design_experiment, draft_interview, capture_evidence, read_analytics, evidence_review. Routes via parent founder skill. Trigger on: "validate my idea", "design an experiment", "Mom Test", "interview script", "capture evidence", "what do we know", "assumption ledger", "read my analytics", "evidence review".

Founder Validation (Phase 2)

Child of founder. Validation support — owns the assumption ledger, experiment design, interview scripts (Mom Test protocol), browser MCP analytics reading (Envelope D), and evidence capture with user-reported artifacts. The user is the validator. This skill is the infrastructure.

Scope: Validation infrastructure. Designs experiments, drafts scripts, captures user-reported evidence, reads analytics, reviews evidence gaps. Does NOT autonomously validate a business. The user runs the experiments and reports back.

Siblings (parent = founder):

• founder-ideation — Phase 1 — adversarial brainstorm + data grounding
• founder-business-model — Phase 2 — calculator mode unit economics
• founder-sprint — Phase 2 — lean gatekeeper stage machine
• founder-gtm — Phase 3 (deferred) — positioning, distribution, channel selection

---

Inherited hard rules (from parent founder): HR-1 through HR-11 all apply. Key inherited constraints: no valuation/legal/tax advice (HR-1, HR-2), no LLM-generated TAM (HR-3), kill criteria required on all ideas (HR-4), data citations required (HR-5), founder is pre-execution only (HR-6), venture-brief is canonical state (HR-7), intake required (HR-8), physical-world bridge preferred (HR-9), epistemic honesty first-class (HR-10), Reddit privacy (HR-11).

---

Modes

1. design_experiment

Given an assumption from the venture-brief, produce an experiment design.

Input:

mode: "design_experiment"
assumption_id: uuid           # links to venture-brief.assumptions[]
assumption_claim: string      # the claim to test (from assumptions[].claim)
biz_type: enum               # from venture-brief intake
niche: string                # from venture-brief intake
stage: enum                  # from venture-brief intake

Flow:

1. Read venture-brief.yaml — load the assumption, existing experiments, intake context. Refuse if missing or if assumption_id not found.
2. Select method from the methods matrix (see references/experiment-design.md):
   • Interview (Mom Test) — for problem/need assumptions
   • Landing page (fake door) — for demand assumptions
   • Concierge — for solution assumptions
   • Pre-order — for willingness-to-pay assumptions
   • Ad test — for channel/audience assumptions
   • Survey — for preference/demographic assumptions (weakest method — flag this)
3. Produce experiment design:

   experiment:
     id: <uuid>
     assumption_id: <uuid>
     method: <selected>
     hypothesis: string           # "If [assumption] is true, then [observable outcome]"
     success_criteria: string     # quantitative threshold
     kill_criteria: string        # what result would falsify the assumption
     minimum_sample_size: int     # method-dependent minimum
     timeline: string             # realistic, not aspirational
     steps: list[string]          # 5-8 concrete steps the user takes
     tools_needed: list[string]   # what the user needs (Calendly, Typeform, etc.)
     cost_estimate: string        # $0 / <$50 / <$200 / custom
     risks: list[string]          # what could go wrong with this experiment
   

4. Write to venture-brief.experiments[] with status: planned
5. Return to user with the full experiment design and "here's what you do next"

Output: Structured experiment design ready for the user to execute.

2. draft_interview

Produce a Mom Test interview script for a specific assumption.

Input:

mode: "draft_interview"
assumption_id: uuid           # which assumption this interview targets
target_persona: string        # who you're interviewing ("UK accountant, 1-5 person practice")
interview_context: string     # optional — "first interview" / "follow-up after landing page"

Flow:

1. Read venture-brief. Load assumption and any prior interview evidence.
2. Apply the Mom Test protocol (see references/interview-scripts-mom-test.md):
   • Talk about their life, not your idea
   • Ask about the past, not the future
   • Less talk, more listen
   • Never pitch — extract
3. Produce a script with 10-15 questions:

   interview_script:
     target_assumption: string
     target_persona: string
     warm_up: list[string]        # 2-3 context-setting questions
     core_questions: list[
       question: string
       follow_up_prompts: list[string]
       red_flag_answers: list[string]  # answers that signal fake validation
       green_flag_answers: list[string]  # answers that signal real pain
     ]
     closing: list[string]        # 2-3 wrap-up questions including the "ask"
     interviewer_notes:
       - "If they say 'I would definitely use that' — DO NOT count this as validation (HR-V1)"
       - "If they describe a workaround they've built — this is strong behavioral evidence"
       - "If they can't describe the last time they had this problem — the pain may not be real"
   

Output: Complete interview script with red/green flag guidance.

3. capture_evidence

User reports what happened. Skill structures it into the assumption ledger.

Input:

mode: "capture_evidence"
experiment_id: uuid           # links to venture-brief.experiments[]
evidence:
  type: enum                  # interview | landing_page | ad_test | survey | concierge | other
  date: date
  assumption_id: uuid
  method: string              # "Mom Test interview" / "fake door landing page" / etc.
  raw_data:
    interviews_completed: int
    replies_received: int
    objections_heard: list[string]
    prepay_asks_made: int
    conversion_events: int
    qualitative_notes: string

Flow:

1. Read venture-brief. Load the experiment and assumption.
2. Validate the evidence artifact (HR-V5):
   • Reject if raw_data has no quantitative fields filled (interviews_completed, replies_received, conversion_events, prepay_asks_made — at least one must be > 0 or the user must explain why)
   • Reject "it went well" without numbers
   • Accept zero values if the user explicitly reports "0 conversions out of 50 visitors"
3. Determine verdict per assumption:
   • confirmed — ONLY with behavioral evidence (HR-V1): signup, prepay, workaround, switch. Verbal "I'd buy it" does NOT qualify.
   • falsified — experiment ran, results clearly below kill criteria
   • inconclusive — experiment ran, results between success and kill criteria, or sample too small to conclude
4. Assign confidence: high (large sample, clear signal) / medium (adequate sample, some noise) / low (small sample or ambiguous signal)
5. Write to venture-brief:
   • Update experiments[experiment_id].evidence with the artifact
   • Update experiments[experiment_id].verdict
   • Update experiments[experiment_id].status: completed
   • Update assumptions[assumption_id].evidence[] with reference
   • Increment interview_count if type is interview
6. Return structured summary including verdict, confidence, what it means, and what to test next

Verdict logic detail:

Evidence type	Confirmed requires	Falsified requires
Interview	Described workaround, switched from competitor, committed time/money	Could not describe the problem, no workaround, indifferent
Landing page	Signup rate > success_criteria, email collection > threshold	Signup rate < kill_criteria after minimum traffic
Ad test	CTR > success_criteria, CPC < ceiling	CTR < kill_criteria after minimum impressions
Pre-order	Pre-orders > 0 with real payment method	Zero pre-orders after adequate exposure
Concierge	Repeat usage, referral, willingness to pay for continued service	Dropped after first session, no referral
Survey	WARNING: surveys are the weakest method. Never mark confirmed from survey alone.	Can mark falsified if strong negative signal

4. read_analytics

Browser MCP Envelope D. Reads analytics dashboards and pages for evidence capture.

Input:

mode: "read_analytics"
target: string                # URL or description: "my GA4 dashboard", "landing page at example.com"
metrics_of_interest: list[string]  # what to look for: "signups", "bounce rate", "CTR"

Flow:

1. Verify browser MCP availability. If browser MCP tools are not available, return a graceful degradation message: "Browser MCP is not available in this session. Ask the user to provide the analytics data manually."
2. Navigate to the target using ONLY Envelope D tools (see references/browser-mcp-allow-list.md):
   • navigate to load the URL
   • get_page_text to extract visible content
   • read_page for DOM structure (forms, CTAs, key elements)
   • read_network_requests for XHR/Fetch responses (GA4 JSON, GSC JSON)
   • find for natural-language element search
   • javascript_tool ONLY for read-only expressions (regex-gated — see allow-list)
   • computer ONLY for screenshot + scroll — NO clicks, NO typing
3. Extract structured metrics:

   analytics_snapshot:
     url: string
     captured_at: timestamp
     metrics:
       - name: string            # "signup_rate", "bounce_rate", "sessions"
         value: string           # "3.2%", "150", "45s"
         source: string          # "page_text", "network_xhr", "dom_element"
     observations: list[string]  # "Landing page has no clear CTA above the fold"
     raw_data_available: bool    # whether XHR captured structured data
   

4. Return for evidence capture. The analytics snapshot feeds into capture_evidence mode.

Envelope D enforcement: Any tool call not on the allow-list is BLOCKED immediately (HR-V3). The skill does not attempt the call and explains why it was blocked. See references/browser-mcp-allow-list.md for the canonical allow-list with regex patterns for javascript_tool gating.

5. evidence_review

Across all experiments, what do we know? What's still unvalidated? What should we test next?

Input:

mode: "evidence_review"
# No additional input needed — reads everything from venture-brief

Flow:

1. Read venture-brief. Load all assumptions, experiments, and evidence.
2. Build assumption status matrix:

   assumption_matrix:
     - id: uuid
       claim: string
       experiments_run: int
       latest_verdict: enum       # confirmed | falsified | inconclusive | untested
       confidence: enum
       evidence_summary: string   # 1-sentence summary of what we know
       risk_level: enum           # high (core viability) | medium (growth) | low (nice-to-have)
       next_action: string        # "run experiment X" / "sufficient evidence" / "pivot needed"
   

3. Gap analysis:
   • Which high-risk assumptions have zero evidence? (HR-V4 — surface honestly)
   • Which assumptions have only survey data? (flag as weak)
   • Which assumptions have contradictory evidence across experiments?
   • Are there enough interviews logged? (HR-V2 check for sprint readiness)
4. Sprint readiness check: can the venture advance to the Decision stage?
   • Top-3 riskiest assumptions each have >= 1 experiment with recorded evidence

   • = 1 real interview logged (HR-V2)

   • No high-risk viability assumption falsified without pivot or accepted_risk
5. Return structured review with the matrix, gaps, sprint readiness verdict, and recommended next experiments

Output:

evidence_review:
  total_assumptions: int
  tested: int
  confirmed: int
  falsified: int
  inconclusive: int
  untested: int
  interview_count: int
  sprint_ready: bool
  sprint_blockers: list[string]  # what's missing for Evidence -> Decision transition
  gap_analysis: list[string]
  recommended_next: list[string]
  risk_flags: list[string]

---

Evidence Capture Protocol

The skill requires user-reported artifacts to close the loop. Evidence is NOT inferred from analytics alone — the user must explicitly report what happened.

evidence_artifact:
  type: enum           # interview | landing_page | ad_test | survey | concierge | other
  date: date
  assumption_id: uuid  # links to venture-brief assumption
  method: string       # "Mom Test interview" / "fake door landing page" / etc.
  raw_data:            # user-reported
    interviews_completed: int
    replies_received: int
    objections_heard: list[string]
    prepay_asks_made: int
    conversion_events: int
    qualitative_notes: string
  verdict: enum        # confirmed | falsified | inconclusive
  confidence: enum     # high | medium | low
  verdict_rationale: string

See references/evidence-capture-protocol.md for the full protocol including verdict logic, confidence scoring, and edge cases.

---

Cross-Skill Integration

founder (parent)
  | intake populated, route to founder-validation
  v
founder-validation (this skill)
  | read venture-brief
  |
  |-- design_experiment --> writes to venture-brief.experiments[]
  |-- draft_interview --> returns script to user
  |-- capture_evidence --> user reports back, updates venture-brief
  |-- read_analytics --> browser MCP Envelope D, feeds capture_evidence
  |-- evidence_review --> gap analysis, sprint readiness check
  |
  v
founder-sprint (Phase 2)
  | checks evidence_review.sprint_ready before Evidence -> Decision gate

---

Failure Modes

Failure	Detection	Response
No venture-brief / missing intake	File missing or intake incomplete	Return to parent: "intake missing"; parent asks the user
Assumption not found	assumption_id not in venture-brief.assumptions[]	Return error: "assumption {id} not found in venture-brief"
Evidence lacks raw numbers	capture_evidence validation fails (HR-V5)	Reject with: "evidence must include raw numbers — how many interviews? how many conversions?"
User claims "confirmed" on verbal intent only	HR-V1 check fails	Reject verdict, explain why verbal intent is not behavioral evidence
Browser MCP unavailable	Tool call fails or tools not in environment	Degrade gracefully: "Browser MCP not available — please provide analytics data manually"
Browser MCP tool not on allow-list	Tool name not in Envelope D list (HR-V3)	Block immediately, explain: "tool X is not permitted in Envelope D (read-only analytics)"
Zero experiments for an assumption	evidence_review surfaces it	Flag honestly as "untested" (HR-V4), recommend experiment
Schema version mismatch	venture-brief schema_version != 2	Error and refuse to proceed (HR-7)

---

Anti-Patterns

Anti-Pattern	Why It Fails	Correct Approach
Marking "confirmed" because the user said "they loved it"	Verbal intent is the #1 source of false validation; people are polite, not committed	Require behavioral evidence: signup, prepay, workaround, switch (HR-V1)
Skipping interviews and going straight to landing page tests	Landing pages test messaging, not problem existence; you need interviews first	Interview first (Mom Test), then landing page to test positioning
Treating survey responses as strong evidence	Surveys measure stated preference, not revealed preference; gap is massive	Flag surveys as weak; never mark confirmed from survey alone
Running experiments without clear kill criteria	Without kill criteria, every result is "encouraging"	Experiment design MUST include kill criteria before execution
Using browser MCP to fill out forms or click buttons	Violates Envelope D (HR-V3); the skill is read-only	BLOCK any write action; explain the Envelope D boundary
Inferring evidence from analytics without user confirmation	Analytics can be misleading (bot traffic, self-visits, etc.)	Analytics feed into capture_evidence; user confirms the narrative
Advancing to Decision stage with 0 interviews	HR-V2 violation	Refuse and surface the gap: "0 interviews logged, need at least 1"
Fabricating confidence from lack of disconfirmation	Absence of evidence is not evidence of absence	Surface as "untested" or "inconclusive" (HR-V4)

---

Reference Files

Read these as needed during validation work:

• references/assumption-ledger-schema.md — YAML schema for structured assumptions, linking to experiments and evidence
• references/experiment-design.md — methods matrix (interview / landing page / concierge / pre-order / ad test / survey), sample sizes, timelines, success criteria templates, cost estimates
• references/interview-scripts-mom-test.md — Mom Test protocol, question templates per assumption type, red flags, anti-patterns, "the three critical questions"
• references/browser-mcp-allow-list.md — Envelope D tool allow-list with regex patterns for javascript_tool gating, blocked tools with reasons
• references/evidence-capture-protocol.md — how to structure user-reported artifacts, verdict logic per evidence type, confidence scoring, edge cases

---

When NOT to Use This Skill

• User wants to generate ideas (not validate them) -- use founder-ideation
• User wants unit economics / pricing -- use founder-business-model
• User wants to build the product -- hand off to forge via sprint
• User wants legal / regulatory validation -- REFUSED (HR-1, HR-2); refer to counsel
• User wants to scrape competitor sites -- NOT in scope; Envelope D is read-only analytics for the user's own properties
• User wants to automate outreach (Envelope B/C) -- deferred to Phase 2.5; Envelope D is read-only