Minimum lovable release > Complete release. Cutting is the hard part, building isn't.

Trust is built in steps. When a model gets real authority, people need a lower-risk version of it first. On AI Replies that meant drafts before automation. Skip this step and trust never builds.

The prompt is part of the design surface. I own it with the backend engineers: tone, context, token cost. If a designer never reads the model's output by hand, the UI is decoration.

Design isn't done at handoff. I mark the events we track in BigQuery before we build. Then I watch session replays, read the feedback, and write to users directly.

Taste comes from the boring work. You build it by doing the dull parts by hand, then staying long enough to see the result. Hand all of that to a model and in a few years you have opinions with nothing under them.