The eval is the spec
Outputs are tickets shipped. Outcomes are customer behavior that changed.
Counting tickets closed is the oldest PM habit there is. None of those numbers prove a customer behaves differently.
Outcomes thinking means naming the behavior change before anyone builds, and the signal that will prove it. One sentence: this will cause a specific customer to do a specific thing, by how much, by when. If you can't write it with a number and a name, you're not ready to build.
For AI features the eval is that sentence made executable. Define what good means before the code exists: a rubric, twenty test cases across the happy path, the edges, and the adversarial inputs. Wire it into CI and nobody ships a regression without you knowing.
Generating options is close to free now. The scarce half of this skill is selection: killing the ideas that won't move the outcome before they eat a sprint.
Today
Write the outcome sentence for the thing you're building right now. If it's an AI feature, write the first ten eval cases too.
Done when: Someone on your team could read the sentence and tell you, in a month, whether it worked.
Get one lesson a day instead
Same lessons, one a day, so the tasks actually happen.