Workflows

Workflows

Workflow: Validate an Idea

2 of 2 in this book · 2 min read

Inputs

templates/idea-input.md (idea, audience, country, trend, problem, offer, pricing, distribution, assumptions, research budget). Unknown fields are allowed.

Stage 0 — Frame (Orchestrator)

  • Name the single main hypothesis; specify intended buyer vs end user, geography, market, alternatives, desired frequency and measurable success.
  • Identify up to 5 critical assumptions; do not prematurely design features.
  • Create output directory; initialize evidence and claim registries.

Stage 1 — Evidence collection (parallel if available)

  • market-researcher: 1 Demand, 2 Pain, rough sizing with explicit method.
  • trend-analyst: 4 Trend, 5 Behavioral Shift; time horizons and leading indicators.
  • growth-analyst: 6 Emotion, 7 Viral, 8 Retention, 9 Community Effects.
  • business-analyst: 3 WTP, 11 Scalability; pricing, unit economics scenarios.
  • competitor-analyst: 10 Advantage, current alternatives and switching costs.
  • execution-analyst: 12 Validation Cost, falsifiable MVP experiment.

Each agent outputs evidence rows with source URL, title, publication date, access date, population/region and exact supported claim. Prefer primary sources, credible surveys, real competitor sites/terms. At least one counterexample/alternative explanation per major assumption. No duplicate research if evidence already exists.

Stage 2 — Evidence audit

  • Remove duplicate or off-topic sources, reject unverifiable/ambiguous metrics.
  • Cross-check dates and definitions: users vs active users, gross vs net, views vs paying customers, AI correlation vs causation.
  • Mark evidence strong / moderate / weak / unverified, confidence high / medium / low.
  • Mark factual claims without evidence as assumptions. Identify legal, safety or privacy blockers when relevant.

Stage 3 — Scorecard

  • Use rubrics/evaluation-scorecard.md. Enter 0–5 only for supported criteria; blank = unknown. Weights sum to 100.
  • Do NOT backfill unknown with 3/5. Distinguish weighted_observed_score (among scored criteria) and coverage_weight.
  • Explicitly report unknown criteria. Scores are prioritization heuristics, not forecast probabilities.
  • Optionally python scripts/score.py outputs/<slug>/scorecard.csv.

Stage 4 — Red team

  • Challenge why this could fail, AI-as-competitor, acquisition costs, contribution incentives, bias, sampling, false positives, and platform dependence.
  • Prioritize at most 5 dealbreakers; for each propose cheap decisive evidence that could overturn initial recommendation.
  • ONE evidence-repair round ONLY for high-impact gaps with accessible data.

Stage 5 — Decision and next tests

  • Output decision-report.md, research.md, red-team.md, experiments.md, scorecard.csv, evidence-ledger.csv.
  • Decision meanings: GO_TO_EXPERIMENT = worth small test; TEST_MORE = unresolved critical assumptions; PIVOT = change buyer/problem/channel; STOP = decisive negative evidence. Never label desk research as market-proven GO.
  • Each test: hypothesis, target sample, acquisition channel, offer, success metric, threshold set before running, time/cost cap, failure signal, next decision.

Stop conditions / cost guardrails

  • If no browsing access, stop factual claims and produce an evidence acquisition plan.
  • Do not run indefinite agent debates. Max 1 specialist pass + 1 challenge + 1 repair; report unresolved gaps.
  • Avoid researching 10 unrelated ideas in one run; isolate ideas then compare using the same rubric.
  • Do not auto-publish, auto-pay, contact research subjects, or expose CV/private user data.