A site where people praise or roast each day's four NYT Connections categories, seeded by AI so the page is never empty.

• Daily puzzle ingestion: Pulls today's categories from NYT Connections data
• AI bootstrap: LLM rates each category 1-5 as initial seed data (1 = dumpster fire, 5 = chef's kiss)
• Voting system: One vote per IP per puzzle, real-time average updates
• Leaderboards: "Best Days" and "Worst Days" ranked by highest/lowest overall scores
• Mobile-friendly: Clean, responsive interface

• 5 = Chef's kiss 🔥 (brilliant wordplay, fair difficulty, satisfying "aha!" moment)
• 4 = Pretty good ⭐ (solid category, maybe one tricky word)
• 3 = Okay 😐 (fine but unremarkable, or slightly unfair)
• 2 = Rough 😬 (too obscure, unfair difficulty spike, or forced connections)
• 1 = Dumpster fire 🗑️ (completely unfair, impossible without guessing, terrible wordplay)

├── backend/
│   ├── database/
│   │   ├── migrations.ts    # Database schema ✅
│   │   └── queries.ts       # Database operations ✅
│   ├── ai/
│   │   └── rating.ts        # AI rating generation ✅
│   ├── routes/
│   │   ├── ingestion.ts     # NYT data processing ✅
│   │   ├── api.ts          # API endpoints (Phase 3)
│   │   └── static.ts       # Static file serving (Phase 3)
│   └── index.ts            # Main Hono app (Phase 3)
├── frontend/
│   ├── components/
│   │   ├── App.tsx         # Main React app (Phase 4)
│   │   ├── PuzzleList.tsx  # Home page list (Phase 4)
│   │   ├── PuzzleDetail.tsx # Voting interface (Phase 4)
│   │   ├── BestLeaderboard.tsx # Best days leaderboard (Phase 5)
│   │   └── WorstLeaderboard.tsx # Worst days leaderboard (Phase 5)
│   ├── index.html          # Main HTML template (Phase 4)
│   ├── index.tsx           # Frontend entry point (Phase 4)
│   └── style.css           # Custom styles (Phase 4)
├── shared/
│   └── utils.ts            # Shared utilities ✅
├── cron.ts                 # Daily data ingestion ✅
└── README.md

• NYT Connections data: https://raw.githubusercontent.com/Eyefyre/NYT-Connections-Answers/main/connections.json
• 748 puzzles available (as of June 2025)
• Daily cron job ingests new puzzles automatically

• SQLite schema with puzzles, votes, AI seeds tables
• Query functions for CRUD operations
• IP-based vote deduplication
• Average calculation and leaderboard ranking
• Status: v2 schema with category words support

• NYT data fetching and parsing with category words
• OpenAI integration for category rating with full context
• Cron job for daily automation
• Historical backfill on first run
• Status: AI now sees actual words for much better ratings

• Main Hono app with error handling
• Full CRUD API endpoints with category words
• IP-based vote validation and deduplication
• Real-time average updates
• Status: All endpoints include category words

• React 18.2.0 with TailwindCSS
• Hash-based routing without React Router
• Home page with puzzle list and ratings
• Mobile-responsive design
• Status: Clean, fast, functional UI

• Four category cards with NYT colors and words display
• Five-star voting interface with real-time updates
• IP-based vote state management
• Error handling with specific messages
• Status: Full voting experience with category words visible

• Improved AI ratings: AI now sees actual category words for realistic ratings
• Category words display: Users can see the 4 words in each category
• Better rating accuracy: AI evaluates fairness, difficulty, and wordplay quality
• v2 database schema: Clean slate with enhanced data structure
• Status: AI ratings are now much more realistic and helpful

• Database: SQLite with versioned table names for schema changes
• AI: OpenAI GPT-4o-mini for cost-effective rating generation
• Frontend: React 18.2.0 with TailwindCSS for rapid development
• Backend: Hono for lightweight, fast API routes
• Deployment: Val Town for serverless hosting and cron jobs
• IP Hashing: SHA-256 with fixed salt for vote deduplication

• ✅ First visitor each day sees AI seed ratings within seconds
• ✅ Duplicate votes from same IP/puzzle are rejected with clear messages
• ✅ UI is clear, mobile-friendly, and fun—roasting NYT should feel delightful
• ✅ Category words are visible so users can judge the connections themselves
• ✅ AI ratings are realistic and helpful (now sees actual words, not just category names)
• ✅ Leaderboards provide engaging way to discover "best" and "worst" puzzles

• Context-aware: AI now sees the actual 4 words in each category, not just the category name
• Better accuracy: Ratings like [5,3,4,5] reflect actual puzzle difficulty and fairness
• Realistic evaluation: AI can judge wordplay cleverness, connection fairness, and solving satisfaction

• Full transparency: Users see all 4 words in each category (e.g., "BALL • BONE • FRISBEE • STICK")
• Better voting: Users can make informed ratings based on actual puzzle content
• Enhanced UX: Category cards show both the connection name and the specific words

"THINGS A DOG CAN FETCH" = 5/5 (chef's kiss)

• Words: BALL, BONE, FRISBEE, STICK
• AI assessment: Perfect category - obvious, fair, satisfying!

"GREAT MANY" = 3/5 (okay)

• Words: DROVE, HOST, LOAD, SCORE
• AI assessment: Trickier - these can mean "a lot" but not immediately obvious

This represents a major improvement in both AI accuracy and user experience!

• Colored categories: Homepage now shows each category with appropriate NYT colors (yellow, green, blue, purple)
• Dual leaderboards: Replaced "Spiciest Days" with separate "Best Days" and "Worst Days" rankings
• Better navigation: Clear distinction between highest-rated and lowest-rated puzzles
• Improved UX: Category colors match the detail view for consistency

• New API endpoints: /api/leaderboard/best and /api/leaderboard/worst
• Enhanced routing: Support for /best and /worst routes
• Backward compatibility: Legacy /spicy route redirects to best leaderboard

• Fixed vote counting: Voting for all 4 categories now counts as 1 vote (per unique user), not 4 separate votes
• Home page refresh: After voting, returning to home page now shows updated vote counts instead of cached data
• Accurate statistics: Vote counts now properly reflect unique users rather than individual category votes

This project showcases Val Town's capabilities for full-stack development with cron jobs, SQLite storage, and AI integration.