DopaShield

Inspiration

It started with a familiar scene: Sitting down to study at 7 PM, opening my laptop with genuine intent to focus. "I'll just check Instagram for 2 minutes." Three hours later, it's 10 PM, and I haven't written a single sentence of my essay. We've all been there. But this isn't a willpower problem, it's a design problem. Modern platforms are engineered to hijack your dopamine system. During exam season, we watched friends lose entire days to browser rabbit holes, not because the work was hard, but because 5-minute breaks became hour-long spirals. The question: What if AI could catch you before you fall into the loop? That's when DopaShield was born, an AI guardian that protects your most valuable resource -"sustainable focus"

What it does

DopaShield is a browser-based AI agent that detects "cheap dopamine loops" during study sessions and provides gentle, personalized interventions powered by Claude AI. Here's how it works: Pattern Detection (Privacy-First) Monitors behavioral signals, scroll speed, tab-switching frequency, session duration. Detects signatures of dopamine loops (rapid scrolling, jumpy attention, mindless clicking). Crucially: Observes patterns, not content, no tracking of what you're reading or watching. AI-Powered Intervention When a loop is detected, Claude analyzes the behavioral pattern. Generates personalized, context aware interventions. Sustainable Habit Building Tracks focus time saved vs. time lost to loops. Shows progress over days/weeks. Celebrates wins without being preachy. Learns your patterns to give better interventions over time

How we built it

Key Technical Decisions: Local-first processing: Pattern detection happens in the browser to protect privacy Claude API integration: Used streaming responses for real-time feel, prompt caching for efficiency Threshold-based triggers: Only send to Claude when pattern confidence is high (reduce API calls) Non-intrusive UI: Toast notifications, not full-screen blocks, respects user agency. Tech: Chrome Extension + Claude Sonnet 4 API + Node.js backend Architecture: Browser patterns → Local detection → Claude API → Gentle intervention Timeline: 7 hours from idea to working demo.

Challenges we ran into

Mobile reality check - 95% of TikTok/Instagram is on apps, not browsers. Pivoted to "study session protection", where laptops actually dominate.
Defining "dopamine loop" algorithmically - What's a loop vs. legitimate browsing? Built weighted scoring: scroll velocity + tab switches + session patterns.
Privacy vs. effectiveness - Claude wants context; users want privacy. Solution: Pattern data only, {scroll_speed: 520} and {sitename: "youtube.com"} and not site data or browsing history.
Making Claude fast enough - Prompt caching + streaming + pre-generated templates = sub-1-second interventions.
Avoiding "annoying parent" syndrome - Changed tone from "STOP wasting time!" to "Hey, want to reset together?"

Accomplishments that we're proud of

Solves real pain every student faces Privacy-first from day one, no shortcuts Deep Claude integration (streaming, caching, empathetic reasoning) It actually works (caught us in loops during development!) Built ethics + business model alongside the code

What we learned

Technical: Manifest V3 is quirky. Behavioral patterns are noisy. Claude nails empathetic tone. Design: Less is more. Timing matters. User agency is sacred. Product: Scope ruthlessly. The problem sells itself. Validate platform assumptions early. Personal: Hackathons reward focus. Ship, then polish. Clear roles under pressure = win.