The skill that survives AI

I started with a simple claim: when you build a product with AI, project management skills matter more than programming skills. AI writes the code — the bottleneck is knowing what to build, in what order, and when to say "not now."

Then I argued with Claude about it, and the claim kept getting stronger in an unexpected way.

First objection: you still need engineering judgment to review architecture. My answer — no, you need to set the right constraints, and AI can check its own work against them. Even the constraints themselves are mostly a catalog of past mistakes, and AI has read a bigger catalog than any of us.

Second objection: some decisions are one-way doors. Bad architecture at scale, data migrations, breaking API changes. But look closer. Rewriting architecture used to cost years of team work — now it's weeks of iterations plus tests. Data migrations are a solved pattern: expand-contract, dual writes, gradual cutover. AI executes all of it if you state the constraints ("zero downtime, rollback at every step"). Even breaking a public API is routine — companies deprecate old versions constantly, clients migrate, life goes on. Technically, almost every door turned out to be two-way.

But here's what survived the argument. Every "one-way door" that stopped being a technical problem turned into a business problem. Breaking an API is easy to code and hard to afford — the price is paid in customers and trust, and it depends on your market position, not your migration algorithm. Google can say "migrate by March." A startup with three enterprise clients maybe can't.

So the final version isn't "PM beats programming." It's harsher: AI keeps eating each layer. Writing code became reviewing code. Reviewing became setting criteria. Setting criteria is becoming something AI proposes on its own. What's left at the bottom is the thing you can't delegate even in theory: choosing between "break compatibility now and lose X customers" or "carry legacy and pay Y forever" — with incomplete information, and your signature on the outcome.

That's the job. Everything above it is temporary.