ProtoVibing

04 · TEST

Talk to actual humans.

Set the decision rule before you collect a single data point.

WHAT THIS PHASE PRODUCES

You leave with

  • A synthetic-persona rehearsal: flagged weak questions, revised guide, ready for live sessions
  • A pre-registered decision rule and experiment OKR
  • Eight to twelve real interview transcripts you’ll synthesize from

THE MENTAL MODEL

How to think about this phase

Test is where most founders cheat themselves. They show the prototype, the customer is polite, and the founder reads politeness as validation. The whole point of the phase is to set up conditions under which the customer’s real behavior (not their politeness) gives you the answer.

Two disciplines do most of the work. First: write the decision rule before the first interview. “Three of five PMs will paste their next real problem” is a rule. “PMs will love it” is not. Second: interview the past, not the future. Ask what they did last week, not whether they would use a thing.

The Mom Test reflex is to remove every leading question, every yes/no question, and every hypothetical. Before you sit down with a real customer, run your guide through a synthetic-persona dry run. Not as a replacement for real conversations -- those are irreplaceable, and talking to actual humans is the hardest and most important thing a founder can do -- but as a rehearsal that exposes the questions that will stall, the answer patterns that will be misread as signal, and the pivots you are not yet ready to make live. ProtoVibing's agents simulate your named ICP so you can walk in already prepared.

Eight to twelve sessions is the sweet spot for a discovery loop. Fewer and you’re fitting noise. More and you’re procrastinating Synthesize. Aim for the floor, then add a session only if a clear pattern hasn’t emerged.

PRACTICE PROMPTS

Paste these into plain Claude or ChatGPT

Pressure-test the discovery guide against the Mom Test

Here is my draft discovery guide: "<paste yours>"

Apply The Mom Test review. For every question:
- Mark it KEEP, REWRITE, or KILL
- For REWRITE, give the exact replacement that probes past behavior instead of opinion
- For KILL, explain what false signal it would generate
- Add the single most important question I forgot

Then run a 60-second mock with a hostile-but-honest persona. Show me where they stall, dodge, or politely lie.

Write the decision rule before testing

My hypothesis: "<paste>"
My P0: "<paste>"
My planned interviews: "<8–12 sessions, with whom?>"

Write the decision rule we will check after these interviews. The rule must:
- Be specific to a behavior I can observe in the session, not a feeling reported afterward
- Have a numeric threshold (e.g. "3 of 5 do X within 90 seconds")
- Define what evidence would falsify the bet entirely

Then write the abandon-the-idea sentence I would have to send to my advisors if the rule fails. Make me feel that sentence before I run a single session.

Run a synthetic-persona dry run

Play three different personas of my named customer in sequence, each for a 6-minute mock interview using my discovery guide. After each:
- Note the question that caused the most awkward silence
- Quote the answer that would have felt like signal but isn’t
- Tell me what you, as the persona, were trying to avoid saying

After all three, list the 2–3 changes to the guide that would make it most resistant to false-positive politeness.

Paste a prompt into Claude or ChatGPT, then replace the bracketed placeholders with your own work. Free, no signup required.

COMMON TRAPS

What goes wrong in this phase

  • Showing the prototype before asking about current behavior. Once they’ve seen it, you’ve contaminated their answers.
  • Treating “I would use this” as evidence. It isn’t.
  • Letting the decision rule drift after the first three interviews look soft. The whole point was that the rule was set in advance.
  • Recruiting friendly people. They are the worst dataset you can collect.

WHAT GOOD LOOKS LIKE

A worked example

SETUP

From the Create example: 10 PMs from Series-B SaaS, recruited via warm intros.

OUTPUT

After Test: 10 sessions, 10 transcripts, and a pre-registered rule of “6 of 10 paste their next real problem within 90 seconds, and 4 of 10 share the generated outline with a teammate within 24 hours.” Real result: 7 paste, 2 share. Decision rule status: paste passes, share fails. Partial signal, must be re-tested with a tightened share-flow in Loop 2.

Practicing this phase? Run the full loop on the platform with your team.