THE ATLAS · AI-NATIVE DESIGN
Every node is something that already exists here: a verified live pattern, a starter-kit template, or case evidence. Hover to trace relations. Open a note to see why it matters.
FIELD NOTES
LocalMode UI gets the order right: capability gate first, download bar second, budget meters after that. Most on-device AI fails as a promise before it fails as a product. Checking what the device can do is an interface decision, not an engineering detail.
Most streaming UI fakes liveness with shimmer. BoardUI’s web search draws each source as it opens - the motion is the evidence arriving, not decoration over a wait. That is the difference between animation as progress and animation as theater.
assistant-ui’s conversation map is the pattern long threads were missing: one tick per turn, the viewport marked, jump by tapping. Scroll position is not navigation. Every long agent thread I design from here gets a map.
Agent Elements ships the states I used to draw by hand: the clarifying question, the reviewed edit, the plan, the handoff, the recoverable error. When a pattern becomes a component, it stops being a nice-to-have and starts being a floor.
Six assistant-ui primitives now cover the chat skeleton: thread, composer, tool fallback, branching, attachments, message actions. The interesting design work has moved up a level - to approval, evidence and recovery. That is where this atlas concentrates.
A modal that says "Are you sure?" teaches nothing. An approval that names the action, the risk and what changes earns its interruption. Every approval surface in my interface lab now answers those three questions before it asks one.