FG—01 · An opinionated reference
AI needs interface design more than interface design needs AI.
A field guide, not a catalog — 7 chapters, 10 principles, 42 patterns, and live specimens you can poke at, for designing AI products that people trust on the second use, not just the first.
Start where you work
Same guide, four doors. Pick the one nearest your desk.
Prototype
Jump into the ServiceTitan office shell — Atlas, dispatch board, and live AI response patterns.
Designer
Start with the patterns — the anatomy, pitfalls, and live specimens you'll cite in crit.
Product manager
Start with chapter one — it's allowed to talk you out of shipping.
Engineer
Start with the worksheets — the checks that gate what actually ships.
The curriculum
Seven chapters, in reading order.
Strategy first — the interface follows. Chapter one is allowed to talk you out of shipping.
“The most senior design decision in an AI product is made before any interface exists — and “no” is a shippable answer.”
Should This Be AI At All?
- CH—02The Model Is a MaterialDesigners who can’t read an eval are designing blind — nondeterminism is a material property, and you learn materials by testing them.
- CH—03Mental Models & the First Five MinutesPeople walk in with a mental model borrowed from sci-fi, search engines, or a coworker — your onboarding either corrects it or gets crushed by it.
- CH—04Legibility Is the ProductIn a system that might be wrong, showing the work isn’t transparency theater — it’s the primary trust interface.
- CH—05Delegation: Designing the Agentic Org ChartWhen AI acts instead of answers, the design question becomes organizational — who approves, who steers, who owns the consequence.
- CH—06When It’s WrongError handling is not an edge case in AI products — it’s a core loop, and products are judged by their worst answer, not their average one.
- CH—07Memory & the Long RelationshipPersonalization without legibility is surveillance with better UX — a system that remembers must show what it remembers and take corrections.
Assistant · proposed action
Every demo in this guide runs in the page. This one implements the approval-gates pattern: consequence stated in figures, reversibility declared, and the human decision made unmissable.
From Approval gates in the Control category.
The arc of an interaction