Home Projects Playground Writing About

Search

Find anything on the site

Esc

Powered by Pagefind. Start typing to search posts, pages, and team members.

Home Projects Playground Writing About

Search

Find anything on the site

Esc

Powered by Pagefind. Start typing to search posts, pages, and team members.

Protected project

GenUI Framework

Enter the password to view this case study.

01 Overview

GenUI
Framework

Generated UI requires predictable anchors and consistent interaction patterns to translate vehicle context into a safe, glanceable driver companion. This is the framework that defines them.

  • Design Systems
  • Generative UI
  • AI
  • 2026

Role

Product Designer / Prototyper.

Pioneered GenUI from 0 to 1 and started cross-functional governance.

Outcome

Translated blue-sky concepts into a tangible architectural AI generative UI framework.

Defined the interaction patterns, rules, and consistent mental models for generative UI components.

Impact

Created an intelligent North Star immediately adopted across visual and mobile teams, directly informing the near-term AI product vision.

Embedded as the GenUI expert across teams, driving advisory and co-creation of intelligent experiences launching in the next release cycle.

02 Background

From no framework to a framework

Phase 1 (GenUI Concepting) was an open exploration: a wide field of generative vignettes, journey-guide moments, and real data integration prototyping for key research insights. It proved what was possible, but it didn't scale. We needed a framework that defined what was consistent, safe, and trustworthy.

Phase 1 · No framework

Open exploration

  • UI vignettes
  • Interaction mental models prototyping
  • Assess engineering feasibility
  • On road user testing & key research findings
  • Answers "what's possible?"

Phase 2 · With framework

A scalable system

  • Build discipline and rules for AI behaviors
  • Categorize the generative UI types across AI moments
  • A framework and system logic any team can apply
  • Answers "what's consistent & scalable?"

A few highlights from Phase 1

Trip to Tahoe: public demo at CES 2026.
Phase 1 breadth of generative UI vignettes
Phase 1: a wide field of generative UI vignettes.
Phase 1 GenUI concepting explorations
Phase 1: real data prototyping and on road testing.

Key findings & takeaways from Phase I explorations

Key Findings

  1. 01

    Binary Actions & Consistency

    Users prefer simple Yes/No or Accept/Decline actions with a consistent mental model across generated content.

  2. 02

    Self-Explanatory Visuals

    Minimize the constant need for users to ask questions; the UI should explain itself at a glance.

  3. 03

    Voice & Center Screen Reserved for Complexity

    No complex interactions in the cluster screen.

  4. 04

    Motion as a Hierarchy Cue

    Users equate high motion and bigger visuals with high importance.

  5. 05

    Context Awareness

    Users expect the vehicle to know their intent, current mindset, emotional state, and environment.

Key Takeaways

Explore Mediated & Ambient State

Shift lower-risk suggestion interactions from high-profile widget pop-ups to persistent ambient notification toasts or mediated states.

Unify Interaction & Decision Model

Ensure all genUI moments share a consistent interaction pattern with a unified binary action mental model across every state.

Simplify Visuals & Reduce Text Density

Minimize text and shift toward graphical representations.

Anchor AI Home

Establish a stable, predictable AI surface that safeguards driver attention zones.

Phase 1, 2025

Want the full exploration?

See the open-ended concepting, prototyping, and user testing that led here.

Read GenUI Concepting →

03 Interaction Models for Mediated State + AI Home

Steering wheel interaction exploration

This is a direct response to GenUI Concepting's roadmap: rather than defaulting to high-profile widget pop-ups, I explored mediated states and ambient toasts as the home for lower-risk AI suggestions, driven entirely by the steering wheel's four-way capacitive controls.

Explorations of the mediated state pattern

Static explorations of different mediated-state and toast interaction models
Exploring the mediated state: an always-present, low-profile toast that expands only when relevant.

Interaction directions

A few directions for how the steering wheel's four-way controls could drive engagement with these states:

Direction A: visual and interaction explorations for navigating between choices.
Direction B: visual and interaction explorations for navigating between choices.
Direction C: interaction exploration from small toast to full widget.
Direction D: a fourth take on hover, browse, and confirm.

04 AI Behaviors

The dimensions of a generated moment

Reliable AI behavior is the foundation of user trust. This framework establishes structured discipline rules to prevent AI from reinventing new patterns on every response.

Every generated moment can be described by three questions: when, what, and how the AI should surface. Each question has a set of named dimensions that become the shared vocabulary for the whole system.

01

When

Timing · When should AI surface?

  • Proactivity · when can the AI act on its own?
  • Boundaries · when should it stay out of the way?
  • History · what patterns has it learned to act on?
  • Urgency · how urgent is it that the user sees this now?
02

What

Content · What should AI surface?

  • Intent · what is the user trying to accomplish?
  • Density · how much detail does the user actually need?
  • Personalization · how should this adapt to preferences?
  • Context · what is the environmental, vehicle, and user state?
03

How

Placement · How should AI surface?

  • Confidence · how certain is the user goal?
  • Agency · how much control does the user have?
  • Prominence · how much attention does this deserve?

Worked Example

Any driving moment can be classified this way. Take a car arriving home: high confidence and high prominence justify a full-screen "Welcome Home" takeover, but urgency stays low, so the layout shift waits until the car is parked and out of reverse, not fired mid-parking.

05 Architecture

Two ways GenUI shows up

Supportive Layer

  • Lives inside the Base UI and enhances core driving tasks

Takeover Layer

  • Sits on top of the Base UI and takes over for complex or high-confidence moments
Supportive layer (camera, situation awareness, map) and takeover layer (overlay) in the cluster
The supportive layer follows Base UI rules; the takeover layer overlays it for high-confidence moments.

System logic flow diagram

Every input, whether a user query or an environment event, flows through the dimensions to resolve into exactly one of the framework's UI types.

Decision tree mapping user queries and environment events through the dimensions to UI types
The decision logic: dimensions in, a single UI type out.

Logic flow example

Example system logic flow: from user input through dimensions to a resolved UI type
Example flow: an input routed through the system logic to its resolved UI type.

06 GenUI Framework

Build Scalable Framework

The framework resolves to a set of UI variations. Different scenarios are evaluated based on the dimensions of when, what, how.

01/08

Multiple Options

Multiple Options dimensional fingerprint
Multiple Options applied: second example
02/08

Single Option

Single Option dimensional fingerprint
Single Option applied: first example
Single Option applied: second example
03/08

Info Display

Info Display dimensional fingerprint
Info Display applied example
04/08

Rich Notification

Rich Notification dimensional fingerprint
Rich Notification applied: first example
Rich Notification applied: second example
Rich Notification applied: third example
05/08

Simple Notification

Simple Notification dimensional fingerprint
Simple Notification applied example
06/08

Critical Pop-Up

Critical Pop-Up dimensional fingerprint
Critical Pop-Up applied example
07/08

Voice Only

Voice Only dimensional fingerprint: minimal visual presence
Voice Only applied example
08/08

Layout Generation

Layout Generation dimensional fingerprint
Layout Generation applied: updated example

07 Agentic UI

Showing the agent's thinking

Every agent action is surfaced as a visible step so the driver always knows what the AI is doing and why, keeping humans in the loop without adding cognitive load.

Chain of Thought

Each reasoning step is shown in real time. The driver sees the AI check traffic, verify park hours, or reroute, not just the final result.

Safety consideration

When vehicle speed exceeds a safe threshold, chain-of-thought visibility is progressively reduced, condensed to a single status line at moderate speeds, and eliminated entirely at highway speeds. The AI continues reasoning; only the surface is suppressed to protect driver attention.

Chain of Thought: AI surfaces a car wash suggestion based on context.

Built with vibe coding. Steering wheel capacitive touch: hover to preview an action, up/down to browse options, left/right to interact with widget content, center to confirm.

Chain of Thought: "Traffic +30 min" surfaces a route-optimization step.
Chain of Thought: a sunset query triggers a visible "checking park closure" step.

08 Tasks, Automation & Goals

From individual tasks to standing goals

The framework defines how AI handles multiple concurrent tasks, how they are categorized into distinct types, and how each type surfaces differently across cluster and center screen.

Tasks and History

One example of how a single persistent entry lets the driver browse all ongoing and completed tasks and switch between them at any time, keeping the agent's work visible and navigable without interrupting the drive.

Tasks and History: a single entry for browsing and switching between tasks.

Built with vibe coding. Steering wheel capacitive touch: hover to preview an action, up/down to browse options, left/right to interact with widget content, center to confirm.

Automation and Goals

Tasks fall into two distinct categories, each with different UI models and interaction patterns in cluster and center stack: automations (rules the driver sets, running in the background) and goals (proactive intents the AI pursues on the driver's behalf). Both span one-time, duration-bound, and standing scopes, with explicit rules for when proactive behavior should surface or stay silent.

Type Scope Initiated by Example
Automation One-time User "Remind me to pick up my order after work today."
Duration User "Keep Quiet Start active for two weeks."
Forever User "Remind me to grab my license when I leave."
Goals One-time User "Plan the best route for today's open houses."
Proactive "Mountains ahead - download offline maps?"
Duration User "Optimize for fuel economy this road trip."
Proactive "Heat wave next week - activate heat protection?"
Forever User "Learn more about my new neighborhood."
Proactive "You use Pro Power on Mondays - prioritize battery?"

In the Cabin

The culmination: cluster and center console working together to show tasks, history, and goals in context. Supportive and takeover layers compose a calm, glanceable companion across the whole drive.

In-cabin vision: cluster and console with GenUI and maps
Cluster and console: tasks and goals surfaced across both screens.
In-cabin vision: second view of the unified cabin experience
A unified, platform-wide in-cabin experience.

09 Contextual Copilot

A copilot that anticipates the trip

Tying it together: a contextual copilot that watches the journey and the world, reopening routes and beating closing times to surface the right UI type at the right moment.

Always-on trip monitoring: I-80 reopened, cancel hotel and resume trip
Always-on trip monitoring: "I-80 is reopened." Resume trip or stay on the route.
Intelligent timekeeper: USPS closing in 40 min reorders the stops
Intelligent Timekeeper: "USPS closes in 40 min" reorders the next stops.

10 Third-Party Integrations

Connecting to the world outside the car

The AI reaches into services, apps, and devices to act on the driver's behalf, completing tasks before they even think to ask.

Drive-Through Ordering

Pulling into a McDonald's triggers the AI to surface the driver's usual order, ready to confirm and send ahead.

Third-party ordering: the AI places a McDonald's order at the drive-through.

Built with vibe coding. Steering wheel capacitive touch: hover to preview an action, up/down to browse options, left/right to interact with widget content, center to confirm.

Cross-Device

The AI coordinates between the cabin and the driver's phone so passengers and pickups are part of the picture.

Cross-device: ask Helen where to pick her up, coordinated with the phone
Cross-device coordination: a pickup shared between the cabin and a phone.

11 Impact

An Intelligent North Star

From 0 to 1

  • Turned blue-sky GenUI concepts into a tangible architectural framework and interaction rules

Cross-functional governance

  • Started the cross-team collaboration that streamlined, consolidated, and unified each phase's AI design

Unified intelligent vision

  • Framework immediately adopted across visual and mobile teams, directly shaping the near-term AI product vision

GenUI advisory

  • Embedded as the GenUI expert across teams, driving advisory and co-creation of intelligent experiences launching in the next release cycle

Where it started

Phase 1: GenUI Concepting

The open exploration and user testing that this framework is built on.

Read Phase 1 →
01 Overview
02 Background
03 Interaction Models
04 AI Behaviors
05 Architecture
06 Framework
07 Agentic UI
08 Tasks & Goals
09 Copilot
10 Integrations
11 Impact

Zhuyuan He

Send a message