Building AI Reading Club: Features & Behind the Scenes
Eugene Yan
A developer built an AI reading companion called Dewey that chats with you about the book you're reading, right inside the page. It's a working prototype, not vaporware, and he documented the whole messy build process.
Based on reporting by Eugene Yan — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Eugene Yan didn't just talk about the idea of an AI reading assistant, he shipped one. AiReadingClub.com is a small web app where you read public-domain Gutenberg books alongside an AI companion named Dewey, who can explain a confusing passage, quiz you on what you just read, recap the whole book so far, or look up a character you forgot about. The spark, by his own account, came from a Twitter exchange between Andrej Karpathy and Patrick Collison musing about what AI-augmented reading could look like. Yan decided to actually build it rather than just theorize.
What's more interesting than the feature list is how deliberately restrained the design is. Dewey only appears when you select text or hit a button, then gets out of the way. Past conversations show up as small unobtrusive stickies next to paragraphs instead of gaudy highlights, because Yan wanted the reading pane to stay clean and text-first. That's a real design choice, plenty of AI tools default to cluttering the screen with chat panes and sidebars because it's easier to build that way. Yan explicitly rejected an early three-pane layout, library, text, and AI assistant crammed side by side, because it felt distracting.
The build process itself is the more useful part of this writeup for anyone tinkering with AI-assisted development. Yan used Claude to draft product requirements with the MoSCoW method, then iterated with it on SVG wireframes for desktop and mobile, then had it design a database schema, which Claude initially over-engineered into half a dozen tables before Yan simplified it to a single books table. From there he moved to v0.dev to scaffold the actual UI, and finally into Cursor to wire up the backend, database, and API calls. The stack underneath is unglamorous but sensible: Next.js, a Supabase Postgres database, Railway for deployment, and a mix of Anthropic, OpenAI, and Gemini APIs, with Gemini Flash doing the heavy lifting by default thanks to its long context window and generous free tier.
Yan is upfront that several ambitions didn't make it into V1. Voice input and output, natural-language chapter navigation, and fiction-specific extras like auto-generated family trees or plot timelines are all on the someday list, mostly shelved because getting them right without hallucinating spoilers or misrepresenting the book takes real engineering care, not just a clever prompt.
My take — AI-written commentary, not fact-checked reporting
This is the kind of AI project I actually want to see more of: small, personal, built to solve a real annoyance rather than to chase a funding round. The tell that it's genuine is the restraint, Yan kept saying no to feature creep and cluttered UI even when the tools made it easy to add more. Most 'AI companion' products could learn from a builder who's willing to simplify Claude's own suggestions instead of shipping whatever the model spits out first.
Read more about this at: Eugene Yan
