The AI skill PMs actually need (it isn't prompting)

Every PM wants to know how to 'do AI'. The honest answer is that the most valuable thing you can learn has almost nothing to do with clever prompts.

Define good, then measure it

AI products fail quietly. A model that's right 85% of the time feels magical in a demo and frustrating in production. Your job is to define, precisely, what a correct answer looks like—and to build a way to measure it at scale.

This is called an eval. If you can write a good one, engineers will fight to work with you, because you've turned a fuzzy goal into a measurable target.

Start small

Take twenty real examples. Label the ideal output for each. Now you have a tiny eval set. Run your model against it every time you change the prompt or the model. You'll immediately see whether you're improving or just moving the failures around.

The PMs who thrive in the AI era aren't the ones with the flashiest prompts. They're the ones who can tell, with evidence, whether the product is actually getting better.