Minerva Next
2025
I joined a two-person engineering team partway through the build. My job was to take an app that only existed as mobile screens and make it ready to grow: I designed the settings and subscription flows, a design system shared by mobile and desktop, and the AI Tutor, a chatbot that can do everything the app does.
Role
Sole UI/UX Designer
Timeline
July – October 2025
Team
1 designer, 2 developers, project owner
Platform
Mobile app, desktop web app
Overview
Minerva helps students learn any subject with AI. Students build flashcard decks in different formats (multiple choice, true/false, fill in the blanks, term and definition) from their own PDFs, links, text, videos and images. The app can generate a deck or students can build their own. As conversational AI spread, the next step was clear: let students do all of this by chatting.
The Impact
25%
Faster handoff
To developers
50+
Design tokens
Across light and dark modes
100+
Product pages
On one component library
Problem
The core features already existed as high-fidelity mobile screens, but there was no documentation and no shared set of components. Each new request, like a desktop version or a settings page, meant redesigning from scratch. On top of that, the team wanted the whole app to work inside a single chat.
01
Undocumented screens
Flows and dependencies lived only in the screens themselves, so every change started with detective work.
02
No shared components
Screens repeated the same elements with small differences, which made desktop and future screens slow to build.
03
A whole app in one chat
Tests, practice, study sessions and planning all had to fit into one conversation without overwhelming students.
Research
With no documentation to rely on, I audited every existing screen to understand its purpose and the flow between screens. For the AI Tutor, I benchmarked how other products handle conversation.
Every screen
Audited for purpose, flow and repeated elements.
3 benchmarks
ChatGPT, Claude and Gemini Notebook.
1 information architecture
Built from the audit and approved before wireframes.
Benchmarking chatbots showed students need guidance, not a blank box.
I studied how ChatGPT, Claude and Gemini Notebook handle conversation, noting what worked and what didn't. The best experiences never left users wondering what to type next.
Key insights
Suggested prompts lower the barrier to starting. A few clear starting points help students begin without writing the perfect question.
Every reply should point to a next step. Replies that end with actions keep students moving toward what they came to do.
Study modes must be easy to tell apart. Options like "think longer" only help when people understand the difference at a glance.
Copy does real work in AI products. The wording decides whether students feel curious and motivated or unsure what the AI can do.
The audit became the information architecture for settings.
Mapping every screen showed which settings the product actually needed. That became an IA for the settings page, which moved into wireframes and high-fidelity designs built on the existing screens and refined through feedback.
Ideation
The biggest debate was about study modes, a feature unique to the AI Tutor that changes how questions are asked. The owner wanted many modes; students had to tell them apart instantly. Defining each mode carefully also gave developers a clear prompt to build for each one.
Early wireframes exploring how study modes and suggested prompts appear in the chat.
01
Fewer, clearer modes
Each mode needed a distinct purpose and a name that explained it, not a long list of similar options.
02
Guided replies
Wireframes tested suggested prompts and action chips after each reply so a conversation never hit a dead end.
03
Motivating copy
Several rounds of wireframe revisions refined the wording to make students excited about what the tutor could do.
Solutions
Six design decisions that made the app ready to scale. Each one solves a problem from the audit or the chatbot research.
Suggested prompts guide students after every AI reply.
Context-aware action chips suggest what to do next, so students move from question to practice without having to think up the next prompt.
Clearly defined study modes change how the tutor asks questions.
Each mode has a distinct purpose that students can understand quickly, and a defined prompt behind it for the developers.
Sources sit beside the chat on every screen size.
A late feedback session showed students needed their documents while chatting. On mobile, tabs switch between chat, sources and materials. On desktop, side panels expand and collapse, like the sidebars in Notion or Google Docs.
A shared design system took the app from mobile to desktop.
While converting screens to desktop, I suggested a design system and rebuilt the mobile screens with it, giving both platforms one component library. Future screens became faster to design, and developers could see exactly which elements repeat.
Subscription copy helps students choose and the business grow.
The pricing page lays out each plan in full so students can decide with confidence, while highlighting the premium AI features that encourage upgrades that benefit the business as well.
Labels, sorting and filters make the deck library easy to search.
I recommended and added sorting, filtering and custom labels, so students can tag questions and cards their own way and find what they need when they need it.
Key outcomes
Ready for more platforms
A mobile-only prototype became a responsive desktop web app with shared components.
No dead ends in the chat
Action-based prompts and distinct study modes kept students practising.
Faster handoff
A tokenized component library for light and dark modes made developer handoff 25% faster.
Clearer decisions
Plan and settings copy helped students choose confidently.
Lessons
A chatbot can do almost anything, which is exactly why students need guidance. Suggested prompts, action chips and clear modes turned an open-ended tool into one that leads somewhere. Joining midway also taught me that structure pays off: the design system paid for itself as soon as we moved to desktop.
What I’d tell my future self
Guide open-ended AI with clear controls.
A blank text box isn't enough; prompts, actions and modes lead users to value.
Build structure before scaling.
Turning existing screens into components early saves time on every new platform.
Push for clarity over quantity.
More options only help when people can tell them apart.
