Most early-stage builders skip discovery entirely — not because they're lazy, but because AI tools make it dangerously easy to jump straight to building. ChatGPT gives you a confident business plan in 30 seconds. Cursor turns your idea into code in an afternoon. The tooling actively removes friction from the wrong step.
ThinkFirst puts the friction back — intentionally. It intercepts the builder before they open a build tool and asks the questions a good PM would ask: Is the problem real? Who exactly feels it? What have they already tried? How feasible is it technically? Only when the answers are strong enough does it produce a decision-grade output — not a pep talk.
Primary: First-time builders — non-technical, uses AI tools, no PM background, fails at traction stage because they built something nobody needed.
Priority: Once-burned builders — has shipped something before, got no traction, and is actively looking for a better process before doing it again. Highest willingness to pay.
Excluded: Experienced PMs and developers with validated mental models. Developers building purely for themselves. They don't feel the problem.
ThinkFirst runs a Socratic AI conversation — one question at a time, max 20 words per question, no business plan generation, no agreeable AI cheerleading. It works across 5 validation dimensions: user specificity, problem reality, frequency and urgency, existing alternatives, and technical feasibility.
The core mechanic is a finite state machine that separates LLM language generation from backend flow control. Each turn, the model outputs structured JSON — score, next action, response — and the state machine decides what happens next: continue probing, move on after 2 failed attempts, or trigger the 3-pass analysis pipeline. The builder cannot skip ahead. The output is earned, not generated on demand.
The system has a 2-strike recovery pattern — if the builder gives vague answers, ThinkFirst re-asks once from a different angle, then moves on and flags the gap honestly in the output. Hard cap: 10 turns. Safe failure over false confidence.
When the interview ends, the builder clicks Generate Analysis — triggering a 3-pass pipeline that extracts assumptions, scores them across 5 dimensions, and produces a decision-grade validation brief with a clear verdict: Proceed, Investigate, Park, or Pivot.
| Phase | Artifact |
|---|---|
| Problem Discovery | Product Brief v3 — ICP, pain chain, 7-hypothesis validation roadmap |
| Behavior Design | Behavior Spec v1 — tone rules, answer quality detection system, 2-strike recovery, authenticity rules |
| AI Architecture | Core AI Loop v1 — finite state machine, 5-stage PRPAO loop, 3-pass analysis pipeline |
| Prompt Engineering | Conversation prompt v1 — 5 dimensions, strict JSON schema, authenticity checklist enforced |
| 3-Pass Pipeline | Extraction prompt v1, Scoring prompt v1, Decision Brief prompt v1 — each tested in isolation before wiring |
| JSON Schema | Locked schemas for Assumption Map, Eval Scorecard, and Decision Brief — 3 separate contracts |
| Prototype | Working mobile and desktop prototype — React frontend, Node.js/Express backend, Gemini API |
| Eval Framework | Golden Dataset v1 — 8 builder archetypes, 5 eval dimensions, isolation test suite |
| Landing Page | Live landing page with before/after, flow diagram, and report mockup |