When everyone on your team can build anything, shipping velocity becomes the easiest thing to celebrate.
But here’s the trap: asking an agent costs seconds. Asking a colleague costs minutes. Do that a few times a day and you start using your AI as head of product.
Over a quarter, you end up with a pile of decisions no one made together. Every PR is defensible. Every change makes sense in isolation. And yet the product drifts.
I caught myself doing exactly this: aligning at the PR level instead of earlier. The problem is not when we review. It’s what we review.
PRs show code. PRDs show intent. Neither shows behavior.
I wrote about what I think the fix is: prototype the actual behavior, review it together, throw it away, and rebuild it right.
Link in comments 👇