Discovery and research with AI · AI-Native Product Building

Principle

Discovery is compression with receipts. You take hours of noise — calls, reviews, forum threads, your own notes — and distill what is repeatable, observed, and disconfirmable.

AI is a librarian for this work, not a witness. It can cluster themes, draft follow-up questions, and flag contradictions. It cannot tell you what happened in the room when someone went quiet.

The decision

DEC_006

  Raw (primary)
│  interviews · tickets · clips · notes
▼
AI synthesis (derived, cite sources)
│  themes · contradictions · open questions
▼
Human review (you mark agrees / disagrees)
▼
Updated problem bet + next interview script

Raw inputs stay primary; synthesis is a derived layer you can audit.

Inputs worth feeding the model

Input	Good for	Weak alone
Interview transcript	Quotes, pain language, workarounds	Sample size of one
App store / G2 reviews	Recurring complaints, comparison set	Angry edge cases only
Job posts	What companies pay to solve	Aspirational fluff
Support tickets	Real failure modes	Biased to existing users
Your session notes	What you noticed live	Memory without recording

Always keep source links or timestamps. If the synthesis cannot point back, discard it.

Workflow

Prep — problem one-pager from Chapter 02. Write three questions that could falsify the bet.
Collect — five conversations minimum; record with permission or take live notes in the person’s words.
Synthesize — paste raw text into AI with a strict prompt: extract themes, list direct quotes with speaker labels, flag contradictions, do not invent.
Audit — read the raw again. Strike anything the model overstated. Add what it missed.
Update the bet — who narrowed? workaround confirmed? new falsifier?
Decide — validation experiment (Chapter 04), another interview round, or graveyard.

Prompt pattern (copy and adapt)

You are a research librarian, not a strategist.
Sources: [paste transcripts / notes]

Output:
1. Themes (max 5) — each with 2+ direct quotes and speaker
2. Contradictions between sources
3. Behaviors observed (past tense) vs opinions (future tense)
4. Questions for the next interview
5. What evidence would falsify the problem statement?

Do not invent quotes. Say "insufficient evidence" when thin.

Tooling

NotebookLM for long PDFs and multi-doc chat. Claude / ChatGPT for synthesis passes. Otter or native phone recording for interviews. A single spreadsheet as the source of truth — one row per conversation, columns for signal strength.

Common mistakes

Synthesis before collection — elegant maps of nothing.
Treating AI themes as votes — three similar hallucinations are still zero humans.
Only interviewing friends who want to please you.
Losing provenance — pretty summary decks with no way to verify a claim.

Artifacts

templates/interview-synthesis.md — one page per research cycle.
Living evidence table: source · date · behavior observed · quote · strength (weak/medium/strong).