How 100 conversations become 1 non-obvious thesis.

01

The conversation

Each call is 25 minutes, structured in 5 sections (open, background, pain, future, close). Conducted over Google Meet for English-speakers, WhatsApp video for LATAM SMB founders who prefer it. Recorded with verbal consent. Transcribed with tl;dv or Fathom.

• ▸ Open (2 min) — set context, get explicit recording permission, frame as "no sale today."
• ▸ Background (5 min) — team size, message volume, current stack.
• ▸ Pain (10 min) — past behavior questions, never future intent. Mom Test discipline.
• ▸ Future (5 min) — magic-wand framing, willingness-to-pay anchored on a single task.
• ▸ Close (3 min) — clarify, ask for referrals, optional pilot mention.

02

The principles

Borrowed from Rob Fitzpatrick's Mom Test, Steve Blank's customer development, and Eric Migicovsky's YC user-interview canon. Updated for AI-era B2B.

• ▸ Past behavior beats future intent. Always ask "tell me about the last time."
• ▸ Specifics or it didn't happen. "Walk me through your last Monday hour by hour."
• ▸ AI-prep, human-validate. Use Claude/Apollo to source and brief. Never on the call.
• ▸ WhatsApp is the channel and the medium. Book the call inside WhatsApp when possible.
• ▸ Confianza beats funnel. Warm intro converts ~5–10× a cold blast.

03

The capture stack

Markdown files in a private directory. One file per call. Frontmatter holds the structured fields; body holds the prose summary. Files never push to the public repo — the privacy firewall is non-negotiable.

• ▸ Filename: YYYY-MM-DD_vertical_NN.md
• ▸ Frontmatter: vertical, country/city, employees, WhatsApp volume/day, top pain, current workaround, WTP USD, lost-customers/month, magic-wand task, would-pilot, permission status.
• ▸ Body: 200–400 word summary, key insights, surprises, anonymized quote.
• ▸ Aggregates: content/patterns.json (manually curated), content/quotes.json (permission-cleared), content/changed-mind.md (contradictions log).

04

The synthesis

Two passes. Pass one within each call (what did this person uniquely teach me). Pass two across all calls (what do they consistently say that contradicts the consensus). The second pass is where non-obvious lives.

• ▸ Daily — within 1h of each call: 5-bullet markdown dump + one-line takeaway.
• ▸ Every 2 days — promote patterns from "early signal" to "confirmed" if mentioned 3+ times across 2+ verticals.
• ▸ Day 8 — mid-sprint review. If patterns aren't crystallizing, re-segment.
• ▸ Day 14 — Claude pass over all transcripts: "What do these founders consistently say that contradicts the common LATAM-SMB-AI narrative?"
• ▸ Day 15 — write the POV doc. 1 page. 600–800 words. Single contrarian-claim headline. Publishes May 17.

05

The privacy firewall

Two layers in one repository. Private capture (gitignored, never pushed) versus public output (anonymized, permission-cleared, published).

• ▸ Full transcripts and structured records stay local in content/interviews/.
• ▸ Public surfaces (diary, patterns, quotes) only carry data that has cleared the firewall.
• ▸ Quotes attribute to "Owner, [vertical], [city], [N] staff." No names. No business names.
• ▸ Quote permission is collected via WhatsApp after the call, in writing, before the quote ships.
• ▸ Every founder receives the synthesized POV doc on May 17 — the gift, not the heist.

06

Why this is published

Methodology in public is the moat. A copyable site with a copyable counter is replicable. A documented synthesis discipline that turns 100 conversations into a non-obvious thesis is not. Founders who skip the method skip the only durable advantage.

Source

The discipline borrows from The Mom Test (Rob Fitzpatrick), Steve Blank's customer development canon, YC's Talk to Users guide, and Eric Migicovsky's user-interview pattern. Adapted for LATAM SMBs and AI-era B2B by direct contact with the founders this is built for.

← Back to the experiment