Most AI features get approved after one good demo. Then real users bring inputs nobody tried, and the AI gives answers that sound right but aren't.
Before any screens get built, I test the feature on 20 real examples.
1. Collect 20 real inputs.
2. Write what a good answer must contain, and what it must never do.
3. Set the pass bar before you look.
4. Score every output as pass, fix or fail.
5. Run the same 20 again after every prompt or model change.
On LeanDeck, a prompt change that fixed one case could break another, and nothing told us. Now the test set comes first.
Pinterest bringing in a real local client is such an underrated channel! For the delivery partner app, are you keeping it on the same design system as the customer app, or going more utilitarian with big tap targets for riders on the move?