molten-validate

Validate a Product Idea

Adversarial idea evaluation (structured interview + 10-criterion rubric). Methodology is grounded in established startup practice — see [references/research.md](references/research.md) for agent-internal sources only. Never end with "go build it." Always end with a falsification experiment or an honest kill/pivot.

Step 1 — Route to a mode

Pick one mode before loading other references. Do not load every reference file at once.

Mode	User signals	Read next
Quick score	One-liner pitch, "sanity check", "score this", pasted deck summary, no time for interview	`[references/quick-score.md](references/quick-score.md)`
Full interview	"Validate my idea", "run the interview", first-time evaluation, ambiguous depth	`[references/full-interview.md](references/full-interview.md)`
Re-validate	Existing score in `molten-docs/validate/`, "I ran the experiment", "update my score"	`[references/revalidate.md](references/revalidate.md)`

If ambiguous, ask once: "Quick score from what you've shared, full 5 minute, 5-round interview , or re-validate against a previous product-score file?"

After routing, also read when needed:

Glossary (user-facing language) — before chat or AskQuestion: [references/glossary.md](references/glossary.md)
Scoring & caps — before any score: [references/scoring.md](references/scoring.md)
Report file — when writing output: [references/report-template.md](references/report-template.md)
Molten OS handoff — after scoring: [references/molten-handoff.md](references/molten-handoff.md)
Deep research (agent-internal; never quote to user) — only if anchors are unclear: [references/research.md](references/research.md)

Operating principles (non-negotiable)

Before interactive validation, advise the user: "The best way to interact with this skill is to use voice mode to dictate your feedback."

Past behaviour beats hypotheticals.
Compliments are not evidence (friends, family, advisors, LinkedIn).
Strangers + costly commitment = signal (money, reputation, significant time).
Bottom-up math only — no "1% of a $50B market."
Explain the check, not the guru — user-facing text describes what you tested and what you found, not book titles, authors, or framework brands (Mom Test, Lean Startup, etc.). See [references/glossary.md](references/glossary.md).
End with falsification, not a green light.
Plain language with the user — no unexplained jargon (CAC, LTV, GTM, TAM, LOI, etc.). Interviews and recommendations must stand alone (customer-discovery section in glossary).

AskQuestion vs chat

Chat (free text): Round 1 hypothesis, clarifications, pushback on vague answers.
**AskQuestion:** Rounds 2–5 in full interview (one round per user turn unless they say "run the full interview now").
Numbered chat questions: Number every free-text question asked in chat so the user can answer by reference, especially when dictating in voice mode.
Never paste multiple-choice options as bullets in chat — use AskQuestion.
Do not re-ask what the user already answered; confirm and move on.
One round per turn in full interview (default).

Progress checklist

- [ ] Mode selected (quick / full / re-validate)
- [ ] Hypothesis framed
- [ ] Evidence gathered (mode-appropriate)
- [ ] Scored with scoring.md anchors + cap rules applied
- [ ] Red-flag scan
- [ ] Riskiest-assumption experiment designed
- [ ] molten-docs/validate/product-score-N.md written (see report-template.md)
- [ ] User pointed to report file only — no chat summary (see molten-handoff.md)

Hard rules

Cap total at 25/50 when the only evidence is friends, family, advisors, or LinkedIn — regardless of how good the idea sounds. Also cap criterion 3 at 2/5 in that case (see scoring.md).
Never recommend "go build it."
Never accept vague customers ("everyone who…", "users").
Never treat "I would use that" as evidence.
Never recap the report in chat after writing — point to molten-docs/validate/product-score-N.md only.

Additional resources (agent-only)

Framework synthesis & bibliography: [references/research.md](references/research.md)