Research & Synthesis Prompts — Provenance

This page documents how the CurioQuest research base was produced — the methodology behind the four reports (chatgpt-research, claude-research, gemini-research, perplexity-research) and the Research Synthesis that distilled them. Kept so the validation chain is auditable: every conclusion traces back to a known brief, not an unstated one.

The method (why it's credible)

1. One brief, four independent witnesses. The identical Research Prompt was run in parallel across four frontier systems (ChatGPT, Claude, Gemini, Perplexity). Independent agreement → high confidence; divergence → adjudicate, don't average.
2. Skeptical-CSO synthesis. The Synthesis Prompt then triangulated the four into one decision-grade view, explicitly reconciling conflicting numbers and grading evidence.
3. Adversarial framing throughout. Both prompts demand the analyst try to invalidate the idea — "a well-evidenced 'this is a bad business because X' is a successful outcome." No cheerleading.

Research Prompt — key design choices

• Role: "skeptical, independent market and strategy analyst… validate or invalidate, not cheerlead."
• Scope: United States first; prefer sources ≤24 months; flag stale data.
• Evidence discipline (mandatory): every material claim labeled [FACT] / [ESTIMATE] / [ASSUMPTION]; decision-critical stats cross-verified against ≥2 independent sources; "if data is unavailable, say so — never fabricate."
• Fixed output skeleton A–J: exec summary · confidence + top assumptions · Go/No-Go + 3–5 conditions · bull vs bear · TAM/SAM/SOM · competitor matrix · differentiation/whitespace · unit economics · ranked risks · recommended wedge · sources. (This is why all four reports share a comparable structure.)
• The product description embedded here is the as-pitched CurioQuest (freemium "first digital book free," Shopify+POD print tiers $25/$40/$75, subscription, claimed moats incl. owned IP + learning-outcome data flywheel) — and the prompt explicitly says: "The founder believes this is a new product class. Test that claim."

Synthesis Prompt — key design choices

• Role: "skeptical Chief Strategy Officer… triangulate, resolving conflicts and exposing what no single report saw."
• Correlated-bias check (the signature move): the four are NOT truly independent (shared training data, same brief, same US-only scope, likely shared vendor stats) — so the synthesis must ask "where might all four be wrong in the same direction?" and steelman the under-weighted optimistic case.
• Reconcile, don't average: present the range for each key figure, pick the most credible value, show the reasoning.
• 19-section output, notably: consensus-vs-divergence map · reconciled verdict/confidence · moat defensibility ranking · quantified unit-economics model (conservative/base/optimistic + sensitivity) · validation plan with pass/fail kill-criteria + concierge-pilot design · keep/change/kill diff against the Project Objective.

Why this matters downstream

• The synthesis's kill-criteria + concierge-pilot design (section 13) is the direct ancestor of the current pilot-first launch approach and the locked pilot success thresholds (preview→paid ≥6%, repeat intent ≥40%).
• The keep/change/kill diff (section 18) is what drove the Project Objective from the as-pitched freemium model to v0.2 (preview→purchase, owned/human-authored IP, no-photo default).

Connections

• Produced → chatgpt-research, claude-research, gemini-research, perplexity-research
• Distilled into → Research Synthesis
• Drove → Overview, ADR-001, ADR-002, ADR-004
• Raw: Research-Prompt.md · Synthesis-Prompt.md