Gemini in Android Auto was envisioned as a seamless, always-available AI assistant for people on the move. It would need to integrate deeply into the in-car environment while maintaining Google’s standards for safety, reliability, and usefulness.
The project was fast-paced and confidential, involving only a handful of contributors. This limited access demanded precise, high-impact design decisions made with minimal iteration time.
Gemini in Android Auto was introduced at Google I/O 2025.
My Roles
· Conversational AI design
· System prompt design
· Personality engineering
· Content design
Team
- · Conversational AI Designer
- · Product Manager
- · Engineer
Years
- 2023-2024
Contact
email@domain.com
000-000-000
Design Considerations
Gemini In-Dash was built for an in-car environment, and it demanded a fundamentally different design approach from phone or desktop assistants. Every design decision had to balance utility, safety, and cognitive load in a context where attention is scarce and distraction can be dangerous.
Key Assumptions
- - Users would rely on audio and voice as their primary mode of communication.
- - Users could not consistently view or interact with a screen while driving.
- - Safety, not engagement, was the measure of success.
Environment
Driving is dynamic and includes constant motion, background noise, and variable conditions. The system needed to perform reliably whether the driver was in traffic, at high speed, or in a parking lot.
Design focus: resilience under environmental stress—clear communication regardless of noise or motion.
Attention
A driver’s cognitive attention is dominated by the task of driving. The assistant’s interactions could not demand sustained focus or require manual engagement. Every response needed to be instantly interpretable and self-contained, with no dependency on visual confirmation or follow-up taps.
Design focus: minimize cognitive load; use natural language cues to confirm and complete tasks quickly.
Speed
In-car interactions happen in motion. Gemini had to respond quickly enough to be useful in dynamic contexts, like navigation or hazard avoidance, without overwhelming the user with rushed speech. This required balancing response latency with the cadence of natural conversation.
Design focus: tight feedback loops; latency management; prioritizing critical information first.
Clarity
In high-attention environments, ambiguity equals risk. The assistant’s language needed to be short, unambiguous, and directive, avoiding technical or verbose phrasing. Responses had to be structured for immediate comprehension.
Design focus: command-level precision in phrasing, consistent sentence structure, and predictable interaction patterns.
Audio-first
Unlike screen-based assistants, Gemini in Android Auto was voice-native. Voice was both the input and output channel. This shifted the design emphasis from visual UI to auditory UX – the rhythm, tone, and brevity of dialogue. Every design decision had to consider how sound conveyed intent, feedback, and confidence without visual reinforcement.
Design focus: conversational pacing, response length, brevity, and clear verbal affordances that replaced visual cues.
Key Takeaway
In an in-car AI experience, clarity is safety. Every word, pause, and tone must serve one purpose: helping the driver achieve their goal without distraction.
System Prompts
Designing the system prompt…backwards.
Challenge: Many, if not most, conversational AI design processes begin by drafting the system prompt first. However, that approach often leads to overly broad or vague instructions.
Approach:
I flipped the process, and started by writing golden conversations first instead of the prompt. These conversations modeled ideal interactions between the user and Gemini.
- As I wrote, I annotated each turn to identify where the AI needed specific guidance, constraints, or behavioral rules.
- These annotations then informed the structure and tone of the final system prompt.
Outcome:
- The resulting system prompt was grounded in real interaction patterns, not theory.
- It clearly reflected the nuances of in-car use: brevity, precision, and safety.
Key Takeaway: Designing the system prompt from real examples outward creates more natural, context-aware AI behavior.
Conversation Style
Drivers would need access to real-time information while keeping their eyes on the road and hands on the wheel. In this environment, long or unclear conversations aren’t just inconvenient, they’re unsafe and put the driver and their passengers at risk.
Interaction patterns that mimicked mobile or smart speaker use cases like extended back-and-forth dialogues, verbose responses, and unnecessary confirmation turns would not work. In fact, they would pose more of a threat to the user. These patterns risked cognitive overload and prolonged attention shifts away from driving.
The challenge was to design conversational experiences that provided value in as few words and turns as possible – delivering clarity, context, and usefulness without distraction.
Process
Understanding the Driving Context
I began by mapping the constraints of in-car interaction: divided attention, ambient noise, and the cost of every additional second of driver focus. Working with a Product Manager, I defined key use cases where conversation added value (e.g., navigation, charging stations, quick information queries, etc.) and identified where it didn’t.
Designing for Cognitive Load
From there, I audited existing Gemini response patterns to identify friction: redundant phrasing, open-ended prompts, or unnecessary conversational branching. Each was rewritten for brevity and single-intent focus.
Key principles established:
- - Limit speech length to under three seconds per response.
- - Avoid open-ended follow-ups unless essential to task completion.
- - Prioritize confirmation through implicit feedback (tone or phrasing) over explicit repetition.
Prototyping Conversational Flows
I created sample conversations for high-frequency tasks – finding a nearby charger, checking traffic, playing music – and tested them against usability heuristics. The goal was to ensure that every turn served a clear user need while minimizing interaction time.
Tuning Tone and Delivery
Given the high-stakes environment, tone was refined for calm authority. Gemini would be informative, but never chatty. It had to sound helpful, not social; precise, not terse. I iterated to align tone guidelines with situational demands (e.g., directions vs. recommendations).
Solution
The final conversation framework optimized Gemini’s in-car dialog model for brevity, precision, and safety.
Each exchange was designed to:
- - Communicate intent and result within a single turn.
- - Reduce speech duration.
- - Maintain continuity and context without requiring user re-engagement.
Impact
- Safer interactions: Reduced need for multi-turn dialogue lowered driver distraction risk.
- Improved comprehension: Shorter, more predictable utterances increased response clarity.
- Scalable design pattern: Framework and tone principles that could be used for future iterations.
This work turned Gemini’s in-car assistant from a conversational companion into a focused co-pilot, one that helps users stay informed, efficient, and safe while on the move.
Overall Impact
Delivered a communication framework optimized for audio-first, safety-critical environments.
Ensured clarity and responsiveness despite limited visual feedback.
Established a repeatable process for creating system prompts driven by actual conversational evidence, not speculation.
