When building gets cheap, judgment becomes the job

I killed my last product in May. A services marketplace for nomads - booking, payments, providers. It did not survive my own audit: two-sided, so it needed both sides from day one, a cold start that kills most marketplaces, and the only paying demand I could find was by analogy, not evidence.

Inside that dead product was a small feature I never took seriously - a public profile page with a map of the countries you had been to. I threw it out with everything else.

Then I spent a month looking for the next thing. Around 350 ideas, nine runs with AI agents, fifteen named candidates, two finalists that tied and neither pulled ahead. And in almost every run, one motif kept coming back: the movement map. The feature I already had in my hands and had not seen.

That taught me two things, and both changed how I work.

First: when you kill an idea, separate the mechanism from the territory. The marketplace deserved to die - the booking, the escrow, the two sides. The travel-identity underneath it was alive the whole time. Kill the broken machine, not the ground it stood on.

Second, and bigger: the criterion that won was not on my spreadsheet. I spent a month scoring ideas on willingness-to-pay, distribution, moat - and the best ones tied. What broke the tie was that I am a user of this product, I have the energy to carry it for years, and I understand how this kind of thing spreads. The market stays the judge. But solo, at the start, founder-fit outweighed every metric I had been counting.

Here is why this generalizes past my own build.

When design and code get cheap - and with Claude Code, Cursor, and real agents, they do - the bottleneck moves. It stops being “can we build it” and becomes “are we building the right thing.” Execution is no longer the risk. Being wrong about the market is the whole risk. AI built a working prototype of MERIDIAN in days, and that speed is exactly what made the real problem obvious: the hard part was never the code.

So every hypothesis - mine and the ones the AI proposes - goes through one grid before it touches the roadmap:

• KNOWN: real evidence. Data, user behavior, a pattern that holds.
• PARTIAL: a signal, not proof.
• ASSUMED: we want it to be true.

Every ASSUMED gets a kill-or-verify test attached. Nothing rides into the build unmarked - the founder’s gut least of all.

One more rule the audit forced on me: single-player value before any viral loop. MERIDIAN has to be worth opening if you are the only person who ever uses it. It builds the map itself from photo metadata, on device, with zero manual input - so it works on the first open, when you have no friends inside yet. The cold start that killed the marketplace cannot physically happen here. Sharing comes after the solo product is worth it, not instead of it. Strava earned the right to be social because the solo run already tracked itself.

Where it is now: 27 screens, a clickable prototype, about to meet real users for the first test. Not launched, not pretending to be.

The AI-native part is not that AI writes the code. It is that when building gets cheap, judgment becomes the job. The grid is how I keep mine honest.