ReadSmart AI

Shows the extension loaded in chrome://extensions with developer mode enabled
ReadSmart AI Sidebar with Article Summary and Q&A. Demonstrates the sidebar interface showing AI-generated summary.
Multi-Tab Summary feature
Summary Results of multi-tab summary

Inspiration

I had many browser tabs open with articles I wanted to read, but "later" never came. I'd skim headlines, feel guilty, and eventually close them without learning. When Chrome announced built-in AI APIs, I realized I could summarize all those tabs at once without sending my reading history to cloud servers. ReadSmart started as a tool to help me get through my reading backlog while keeping everything private and offline.

What it does

ReadSmart is a Chrome extension that summarizes articles using on-device AI. Click any article to get instant summaries in different formats (key points, TL;DR, teaser), ask context-aware questions with AI-generated suggestions, and batch-process multiple tabs at once in the background. Also includes translation support, and export options for text, markdown, or HTML.

Everything runs locally using Chrome's built-in Summarizer, Prompt, and Translation APIs. Works on 30+ platforms including Medium, Wikipedia, Dev.to, and news sites. Completely offline after initial setup. No API keys, no servers, no tracking. Your reading history stays on your device.

How I built it ReadSmart AI

Built with vanilla JavaScript and Chrome Manifest V3, using three Chrome AI APIs: Summarizer (for article summaries), Prompt (for Q&A and question generation), and Translation. The architecture has three parts: a content script that detects articles using platform-specific CSS selectors and injects the UI, a service worker that handles AI operations and multi-tab processing with state persistence via chrome.storage, and a popup for batch summarization with real-time progress tracking.

Challenges I ran into

Content extraction across 30+ platforms meant building platform-specific selector libraries since Medium, Dev.to, Wikipedia, and news sites all structure articles differently.

Multi-tab state management was the hardest part since service workers can shut down anytime. I had to constantly save progress to chrome.storage and handle restoration without losing data. AI-generated suggested questions were initially generic and useless until I rewrote prompts with explicit examples. Also fixed message ID collisions when users clicked rapidly (added counters to timestamps), memory leaks from uncleaned intervals, and 7 XSS vulnerabilities from unescaped AI content.

Accomplishments that I'm proud of

Works completely offline. Context-aware questions are article-specific, not generic. Multi-tab processing survives popup closes. Truly private - no backend. Supports 30+ platforms with clean UI. I use it daily.

What I learned

The AI prompt engineering was interesting. When I asked it to "generate questions," I got useless generic stuff like "What is this about?" But when I gave it specific instructions with examples of what I didn't want, the quality improved a lot. I also learned that building a library of platform-specific selectors worked much better than trying to detect articles generically. I forgot to clear an interval in one place and it leaked memory on every page load - took me a while to find that one. The Chrome AI APIs were surprisingly fast (3-5 seconds) without needing any API keys or rate limits.