Crux

Video Demo

About this project

Crux is a priority-ranking AI agent that argues with you. Most AI to-do tools sort your list and hand it back. Crux does something different: it ranks your priorities, surfaces the one decision that gates everything else, breaks tasks into actionable steps, and defends its reasoning when you push back. What it does: Three ways in, one reasoning path. Type a list, speak it (transcribed with Whisper), or photograph a whiteboard (vision OCR). Everything funnels into the same ranking engine. Ranking with reasoning. Every item gets a position and a specific justification — not just an order. An adversarial debate loop. Push back on the ranking and Crux weighs your argument. If it holds up, Crux concedes and the list visibly reorders (animated). If it doesn't, Crux holds its position and tells you why. It's a sparring partner, not a yes-machine. Capped at 3 rounds. Three tones — Coach, Consultant, Debater. The voice and framing change; the underlying priority logic stays consistent. Per-item synthesis breakdown. Click any item and Crux returns ordered sub-steps with time estimates, a clear starting point, and only the dependencies that genuinely exist. The hard part: the model kept fabricating structure to satisfy vague prompt rules — inventing task dependencies and authority hierarchies that weren't real. I fixed it with a two-part dependency test written into the prompts: a step only depends on another if it consumes that step's output AND literally cannot begin until it's done. Otherwise, the items are independent and Crux says so. Built with: Next.js, TypeScript, Tailwind, Vercel. Groq Llama 3.3 70B for ranking, debate, and breakdown. Groq Whisper for voice, Claude vision for photo OCR. No database, no auth — browser localStorage only. Built solo in 24 hours at QuackHacks 3. Live at usecrux.tech.