How many customer interviews before you start building?

Most answers to this come with a shrug attached. Ten to fifteen per segment. Until you stop hearing new things. Both are reasonable and neither is usable, because a number you can't check is a number you will have met the moment you decide you've met it.

So here's a specific one and where it comes from. Masterji won't let a goal leave VALIDATION until three proofs of customer conversation are accepted, and it counts distinct people rather than distinct evenings. The phase below it needs exactly one proof: a written problem statement and who has the problem.

Why three, when the books say ten

Three isn't the number at which you understand your market. It's the number at which you've demonstrably left your own head, and those are different jobs with different right answers.

Ten to fifteen is a good answer to how many conversations before you trust a pattern. It's a bad answer to what the bar should be, for a reason I only understood after watching it fail. Set a bar of fifteen interviews in front of somebody who has done none and you haven't built a bar, you've built a wall. What comes out the other side isn't a builder who did fifteen interviews. It's a builder who did none and went back to the doc, because the distance to the next checkpoint is further than an evening reaches.

Three sits where it does because it's the smallest count that can still catch the particular lie this stage invites.

The count isn't the bar. The distinctness is

Three conversations can be three conversations with the same willing friend. A row count cannot see that, and it happens constantly — not because anyone sets out to cheat, but because the friend who'll talk to you about your idea is genuinely the easiest person to talk to, and talking to them three times genuinely feels like three conversations.

So the gate counts subjects rather than submissions. File a second proof about somebody already counted and the coach tells you the person already counted can't be counted again. What you still owe doesn't move.

VALIDATION counts people, not evenings.

That's also the honest reason the number gets to be as low as three. Three distinct people is harder than ten conversations, because ten conversations can include the person who already likes you and three distinct people can't.

Every rung adds a kind of evidence, not more of the same

The same shape repeats the whole way up, and the pattern is the point:

• IDEA — one proof. A written problem statement, and who has the problem.
• VALIDATION — three, about three different people.
• BUILD — two, at least one of them a real user touching the thing. Two artifacts can both be links nobody ever opened.
• LAUNCH — three, at least one of them a stranger acting. Three launch events can be three posts, and posting isn't somebody acting.
• TRACTION — no counter at all. It's terminal: one stranger coming back on their own, or paying, is the finish line.

Read down that list and the original question starts to look like the wrong shape of question, a bit like asking how many push-ups. The count is the cheap half. What the count is of is the expensive half, and it's the half a bar you set yourself will quietly relax when you get near it.

What to do with this if you're not using anything

Keep whatever number you were going to use and spend the effort on the constraint instead. Three people you haven't spoken to before, none of whom owes you a favour, each of whom told you about a specific time the problem cost them something. Write the three names down before you start, so you can't count one of them twice without noticing.

Then the part that decides whether any of it worked: don't be the one who says it counted. Hand the list to somebody, or hand it to something that can't be talked round.

What to actually ask is a solved problem and not mine. The Mom Test by Rob Fitzpatrick is the whole answer and it's short. This is only about the bar, which is the part the book can't hold for you.

If you want to watch a gate refuse an advance before you rely on one, the walk-through doesn't need a sign-in.