AI-driven product-idea funnel — 1000 → 50 → 10 → 1 via a 3-category rubric

Codify the validation rubric you already use by hand for business research into an AI scoring system. Ingest a list of raw idea descriptions, score each against three idea-categories, validate the system against your past decisions to calibrate, then rank composite scores. Final pick is a human filter — "what's closest to my domain familiarity" — over the top-10.

One operator running an idea-factory reports compressing ~1000 raw ideas to ~50 worth attention and ~10 worth building, with the top-1 becoming the next product to ship. The leverage is not the scoring itself — it's that the funnel runs in a day, so you can re-run it whenever the market shifts.

1. Articulate the rubric you already use. Write down what you actually weigh when evaluating ideas manually — market size, ICP clarity, AI-native fit, distribution lever, founder-domain match. Three composite categories tends to be the right ceiling.
2. Translate the rubric into a structured skill. One skill file per category with the questions an agent must answer; one orchestrator skill that aggregates scores.
3. Calibrate against past decisions. Feed in 10–20 ideas you have already accepted or rejected. The system should reproduce roughly your historical picks. If it can't, the rubric is fuzzy — tighten it before trusting forward decisions.
4. Ingest the raw idea list. Source from your own backlog, trend scrapes, app-store deltas, or a deep-research run. Aim for ≥500 candidates so the funnel has signal.
5. Run the scoring. Each idea gets a score per category and a composite. Rank, cut to top-50, then top-10.
6. Human pick at the top. Of the top-10, the founder selects by domain familiarity — not by raw score. The framework finds and ranks; the human picks.

• Trusting the ranking before calibration. Skipping the "validate against past decisions" step is the failure mode — without it, the rubric drifts and the top-10 becomes garbage-in-garbage-out. One operator notes the system itself is "complex" and worth a dedicated writeup; treat v1 as calibration-only.