Build around the verb, not the data

I shipped an internal operating system for my CPG brand, NØRSE CØDE, and used it for about ten minutes before I said, out loud,, that it was useless.

It looked fine. Five pages, clean layout, real data flowing through every one of them. Everything a founder running a company on a team of AI agents would want visibility into. I had built exactly what I set out to build.

The problem showed up the moment I tried to use it instead of look at it.

Every page was a mirror. It showed me what was true, but it gave me nothing to do about it. An order flagged as delayed had no button next to it. A content piece waiting for approval had no way to approve it from that screen. An email thread that needed a reply just sat there, read-only, daring me to go find the real tool that could actually reply to it. I had built five windows into the business and zero doors.

That is a specific kind of mistake, and it is worth naming precisely because it is so easy to make and so easy to miss until you actually try to use the thing. When you design a system by starting from the data model, what tables do we have, what does each row look like, you naturally end up with one page per table. It is the path of least resistance for both a human engineer and an AI agent building from a schema. The result reads as complete. It covers everything. And it is completely useless, because a founder does not sit down to admire a database. He sits down to decide something, approve something, kill something, or find out what needs him right now.

So we redesigned it. The rebuild deleted three of the five pages outright, not because the data behind them stopped mattering but because nobody needed a dedicated screen just to look at it. The two pages that survived got rebuilt around a verb instead of a table. Not "here is the content," but "here is what needs your approval, and here is the one tap way to approve it." Not "here is the inbox," but "here is the reply already drafted, waiting for you to send it or kill it."

The test I use now, on anything before it ships: what is the verb on this screen. If the honest answer is "look," it is not done. Every surface I sign off on has to answer with something a person actually does: approve, send, resolve, decide, kill. If a page cannot name its verb in one word, it gets cut or it gets rebuilt.

This is not a new idea. Anyone who has shipped an admin panel that nobody opened twice has learned some version of it. What makes it worth writing down is how often I re-learn it, and how expensive each round is. Running a company on a stack of AI agents does not exempt you from this. If anything it makes the mistake easier to make, because an agent can generate a competent-looking, fully data-complete dashboard in an afternoon, and competent-looking and data-complete is exactly the trap. The agent optimized for the thing it could measure: does the page render the data correctly. Nobody was measuring the thing that actually mattered: does a human ever act on this page, or does he just glance at it and leave.