"Stop guessing which prompt is better. Duel them and let the data decide."
I designed, coded, and deployed PromptDuel within a single 24-hour sprint on Christmas Day 2025.
The Problem
When developing AI Agents, small semantic changes in a prompt can lead to drastically different outputs. Tracking this in spreadsheets is messy. You need a way to blind test these outputs against each other to get clean, unbiased data.
The Solution
PromptDuel solves the "vibe check" problem. It is a lightweight, structured environment to evaluate LLM outputs side-by-side.
Key Features
- ⚖️ Side-by-Side Arena: A clean, split-screen interface for comparing two text outputs (supports Markdown).
- 🫣 Blind Testing Mode: Model names are hidden from voters to ensure unbiased feedback.
- 🔗 Instant Sharing: Generate public, read-only links for clients or team members to cast votes.
- 📊 Analytics Dashboard: Track vote velocity and win rates visually.
- 🔐 Secure: Row Level Security (RLS) via Supabase ensures data integrity.
Tech Stack
I chose a stack focused on speed and reliability:
- Frontend: Next.js 14 (App Router) + Tailwind CSS
- UI Library: Shadcn/UI + Lucide Icons
- Backend/Auth: Supabase (PostgreSQL + RLS)
- Visualization: Recharts