HiveMind

## The Problem

It's 2 AM before a hackathon. Someone finally asks: "So... what are we building?"

That moment kills more teams than bad code ever will. Most hackers spend
the first 6–8 hours just deciding. By the time they commit to an idea, they're exhausted before writing a single line.

For first-time hackers, ideation alone can eat half the event.
They don't lose because they can't build — they lose because they never got started.

## What We Built

HiveMind runs five AI agents in sequence the moment you enter a theme: Researcher → Ideator → Engineer → Critic → Presenter — each one reading the last, taking you from blank page to pitch-ready concept in ~2 minutes.

The part we're most proud of: after ideation, the agents don't move on.
They argue. A live Debate Room shows the Engineer, Researcher, Critic, and Ideator pushing back on each other in real time until a final call is made.

## What We Learned

Prompting is architecture. Streaming changes everything — the same output delivered token-by-token feels like thinking; delivered all at once it feels like a printout. And race conditions in async UIs will humble you fast.

## Biggest Challenge

Gemini 2.5 Flash burns hidden "thinking tokens" inside your output budget — silently truncating every agent mid-sentence with no error thrown. Finding the bug took a bit of time .

## Who It's For

Everyone at their first hackathon. The solo hacker with no team to
bounce ideas off. The student who's tried three times and walked away with nothing because the first 8 hours always collapsed into indecision.

HiveMind won't write your code. But when the clock starts, you'll already
know exactly what you're building — and why it matters, and have a starting template .

FastAPI · Google Gemini 2.5 Flash · Vanilla JS · WebSockets