NotebookLM: A Knowledge Tool That Turns Your Sources into Anything
From November 2025 to May 2026, I worked as a Product Designer on Google NotebookLM: an AI-powered research tool that lets users synthesize information from multiple sources into notes, summaries, and presentations.
This was my first time working in a fully AI-native product environment. As the sole designer across multiple workstreams, I had to move fast — prototyping in Antigravity, scaling design systems with Figma Make, and making design decisions in a workflow where the product itself was changing what "design" meant.
Across three shipped features, I established patterns that balance AI automation with human agency — and explored how designers themselves can work alongside AI without losing craft.
https://notebooklm.google/
MY ROLE
- Product Designer
- Prototyping with Antigravity / Google AI studio
- AI-assisted Design System (Figma Make)
- Cross-platform design: mobile + desktop
GOAL
- Give users meaningful control over AI-generated outputs without adding friction
- Scale NotebookLM from a generation tool into a personal knowledge system
- Establish AI-assisted design workflows that accelerate without sacrificing quality
Case study 01
Slide editing
Problem
Users can generate AI slides in one click, but they can't control the output. Whatever the AI produces is final. Wanting to change a single slide means regenerating the entire deck.
Process
A UXR insight from the Japanese market from Dec 2025 pointed to the core issue: users didn't dislike AI-generated content. They disliked losing control over it.
The design challenge became: how do you make AI generation and human editing coexist in a single flow? The solution was targeted regeneration — letting users issue new instructions for a single slide rather than all-or-nothing.
I used Figma Make to prototype this interaction, generating initial UI from prompts and iterating quickly on the editing mechanic. Engineering would implement, then come back with feedback — edge cases in the component hierarchy, mobile behaviors that didn't translate cleanly. I'd revise in Figma Make and the cycle continued.
Mobile and desktop required different interaction models for the same editing mechanic. That back-and-forth between design and engineering was where most of the real decisions got made.
The design challenge became: how do you make AI generation and human editing coexist in a single flow? The solution was targeted regeneration — letting users issue new instructions for a single slide rather than all-or-nothing.
I used Figma Make to prototype this interaction, generating initial UI from prompts and iterating quickly on the editing mechanic. Engineering would implement, then come back with feedback — edge cases in the component hierarchy, mobile behaviors that didn't translate cleanly. I'd revise in Figma Make and the cycle continued.
Mobile and desktop required different interaction models for the same editing mechanic. That back-and-forth between design and engineering was where most of the real decisions got made.
In just 2 weeks since initial launch, we've had:
Design
A prompt-based input on each slide lets users give instructions per page. Since changes aren't immediate, a new UI indicator marks which slides have been edited. Users can queue multiple edits and regenerate them all at once.
![]()
![]()
![]()
UXR-driven
Targeted regeneration
Cross-platform
AI x human editing
Case study 02
AI tagging system
AI tagging system
Problem
NotebookLM lets users combine multiple sources — PDFs, web pages, videos — into a knowledge base. But as sources accumulate, finding information becomes difficult and organizing them becomes a burden.
Process
Behavioral observation surfaced a telling signal: power users were manually tagging sources with emoji and custom prefixes to stay organized. The design challenge was making AI suggest tags without removing user agency.
The solution was two parallel flows. One followed a Gmail-style labeling system, letting users create and manage tags manually. The other was one-tap AI tagging, where the system auto-classifies all sources based on content analysis.
The manual flow was deceptively complex: string length constraints, sorting logic, edge cases for long or ambiguous tag names. I worked closely with engineers throughout to make sure the manual tagging logic and AI tagging logic stayed consistent — so regardless of which flow a user chose, the underlying taxonomy made sense.
The solution was two parallel flows. One followed a Gmail-style labeling system, letting users create and manage tags manually. The other was one-tap AI tagging, where the system auto-classifies all sources based on content analysis.
The manual flow was deceptively complex: string length constraints, sorting logic, edge cases for long or ambiguous tag names. I worked closely with engineers throughout to make sure the manual tagging logic and AI tagging logic stayed consistent — so regardless of which flow a user chose, the underlying taxonomy made sense.
Outcome
- Reduced the friction of manual organization to near zero
-
Shifted NotebookLM's positioning from AI tool to personal knowledge system
- Evolved into one-tap AI auto-classification of all resources
Design
Mobile AI tagging: Two creation flows![]()
Mobile AI tagging: Edit, management flow
![]()
Desktop AI tagging, entry point
![]()
Mobile AI tagging: Edit, management flow
Desktop AI tagging, entry point
AI suggestion + human control
Behavioral research
Scalable system design
Case study 03
Optimistic design system
Optimistic design system
AI-built mobile dark mode module
Problem
NotebookLM mobile needed dark mode support, but the design system wasn't complete. Building it from scratch is time-intensive — mobile dark mode requires careful attention to tokens, contrast ratios, and accessibility across every component.
Process
The workflow: AI generates → designer audits → iterate on failures.
I started with our existing light mode tokens, exported them as JSON using Token Studio plug-in, and used Claude to generate a dark mode version based on rules I defined — things like contrast ratios, surface elevation hierarchy, and which brand colors needed to shift in luminosity.
The real work was cleaning up token naming, manually reconnecting nested components that didn't inherit correctly, and auditing every output for accessibility.
Error states defaulted to red text that failed dark mode contrast ratios. Our primary button ︎ (
AI handles generation well. What it can't do is know your system, catch hierarchy errors in deeply nested components, or make judgment calls on brand vs. accessibility tradeoffs. That's still the designer's job.
I started with our existing light mode tokens, exported them as JSON using Token Studio plug-in, and used Claude to generate a dark mode version based on rules I defined — things like contrast ratios, surface elevation hierarchy, and which brand colors needed to shift in luminosity.
The real work was cleaning up token naming, manually reconnecting nested components that didn't inherit correctly, and auditing every output for accessibility.
Error states defaulted to red text that failed dark mode contrast ratios. Our primary button ︎ (
#4259FF) frequently didn't pass WCAG, I resolved it by selecting a lighter blue already in our design system for dark mode use.AI handles generation well. What it can't do is know your system, catch hierarchy errors in deeply nested components, or make judgment calls on brand vs. accessibility tradeoffs. That's still the designer's job.
Outcome
- Independently shipped 300+ dark mode components across mobile and web in under a week
- Established the team's first AI-assisted design system workflow — documented and adopted as the standard process for future work
- Demonstrated that AI-generated output is viable at scale when paired with structured human oversight
Design systems
AI-assisted workflow
Mobile · Dark mode
Accessibilty