An AI reading tool designed to feel like a rabbit hole, not a search result
Project
Curiosity Engine
Focus
learn UX for AI by building something I’d actually use
What I did
Built and designed ... with React, Vite, and the Anthropic API
Context
2026 • Concept
Overview
Inspired by the experience of moving through a well-edited magazine. Following curiosity without knowing where it will lead. Built solo to explore what it means to design for generative output, not just traditional interfaces.
Challenge
Create something that feels like going down a rabbit hole without making the user dig it.
Autonomy without effort, guided enough to feel effortless and open enough to feel like the user is the one exploring.
Competitive Landscape
AI interfaces I learned from
No playbook exists for this kind of interface. But a few products solved specific problems worth stealing from.
Perplexity
Does exactly what it promises. AI-generated answers feel unverifiable by default. Surfacing sources changes that.
NotebookLM
A different use case entirely. Your documents, not the web. The audio overviews are the interesting part: generative AI doesn't have to mean chat.
Claude / ChatGPT
Most useful when you already know what you want. The real lesson came from Copilot: suggestions in a sidebar get ignored. Ghost text that appears where you're already looking gets used.
Curiosity Engine
Draws from all three: sourcing at the content level, reading as the format, and navigation that lives inside the text rather than around it.
Prompt Engineering
Adjusting the prompt until the output was worth following.
The following is problems I observed in early sessions and how I addressed them.
Tone
Card copy that hedged instead of landed
Early copy was written like an essayist, not a person. Headlines sounded meaningful but said nothing. Body copy qualified before it got to the point.
Wrong
"Sleep is essential for memory consolidation, and researchers continue to explore the complex mechanisms involved in how the brain processes information during rest cycles."
Right
"Your brain chemically paralyzes your whole body every night so you don't act out your dreams."
the fix
Made every card lead with the punchline. For each fact, the question became: what's the version of this that sounds like a conspiracy theory but is completely true?
HALLUCINATION RISK
The app could confidently lie about real people
Post-generation verification failed because the model just confirms its own confabulations. If the interface lies with confidence, there's no recovering that trust in the session.
the fix
Before generating anything, a fast check asks the model what it doesnt know. If it can't answer reliably, it asks the user for context first rather than filling the gap with something plausible-sounding.
concept depth
Highlighting a concept often led somewhere adjacent
When users highlighted concepts like developmental trauma, the next card often pivoted to a person, event, or origin story instead of explaining the concept itself. The interaction was working, but the response wasn't matching the user's intent.
the fix
Rewrote the prompt to distinguish between concepts and named entities. Concepts now expand into the phenomenon itself, while specific people, places, and events zoom into the highlighted subject. The AI was also explicitly instructed to avoid defaulting to origin stories or the person who coined a term.
Interaction Design
Interactions for a generative interface - beyond chat
When content is dynamic, interaction must dissolve into the user's existing behavior. The reader becomes a curator by following attention, not navigating an interface.
That flexibility is an opportunity generative content opens.
OUTPUT FORMAT
The output had to feel like a rabbit hole, not an answer.
Every element is designed to open something, not close it. The headline hooks, the body gives you three things to pull on, and the chips point somewhere you haven't been yet.
Reflection