Synthetic Customer Discovery

Synthetic Customer Discovery

A Claude skill for customer discovery: plan real interview studies, build grounded synthetic personas, interview them, and synthesize what you learn into evidence-backed reports.

Built to be honest about what synthetic research can and cannot tell you.

**No API keys. No backend. No accounts.** Plain markdown files that run inside Claude. You can read, version, and correct every persona, and nothing leaves your Claude session.

The problem with synthetic users

Ask an LLM to "act like one of my customers" and it collapses to the average internet idea of one: agreeable, articulate, and suspiciously enthusiastic about your product. Teams run twenty of these "interviews," hear what they hoped to hear, and call it validation.

That isn't research. It's confident fiction.

The fix isn't better prompting. It's structure: personas deep enough that the model has something real to draw on, interview rules that force friction instead of flattery, and outputs that tell you exactly which answers to trust and which to verify with a human.

How this skill keeps you honest

1. Personas are grounded artifacts, not prompt strings. Each one is a structured file built from real episodes from real people you have access to: yourself, your customers, support tickets, sales calls. Every unverified claim is marked [ASSUMED: confirm].
2. Hypothesis, never verdict. Every synthetic output is labeled as a hypothesis. The synthesis route hard-codes it: a theme supported only by synthetic transcripts can never be a finding, no matter how many runs agree.
3. Anti-collapse interview rules. Specificity over archetype. No composite wisdom. Friction, not gratitude. If the persona wouldn't trust your product, the interview has to show it.
4. A mandatory fidelity note on every run. Which answers trace to real grounded material, which are speculation, and the 3-5 specific things a real interview should confirm or kill.
5. **A va