FastAPI · React · GPT-5.6 Sol · Fly.io

PromptGym

A timed arena for AI-assisted building. You get a messy brief and a coding agent, then prompt it to ship a working app before the clock runs out, and a judge LLM grades the result and your prompting.

How It Works

Brief in. Graded build out.

01

BRIEF

A messy, real brief

Every challenge is a garbled meeting transcript: conflicting requirements, half-remembered thresholds, people dropping off the call. You dig out what actually has to get built.

Ambiguous specsImplied requirementsTimed
02

BUILD

Prompt a coding agent

You don't write the code. You direct a GPT-5.6 Sol build agent through a three-column IDE. It writes files and runs commands in a live sandbox; a preview iframe renders the app as it takes shape.

Agentic loopLive sandboxWebSocket stream
03

JUDGE

A judge LLM grades you

On submit, a judge model scores the build across requirements, correctness, code quality, taste, and prompting skill, plus deterministic token-cost and speed percentiles.

5-axis rubricCost percentileTranscript-graded

Stack

FastAPI

Backend

React + Vite

Frontend

GPT-5.6 Sol

Agent + Judge

WebSockets

Live stream

subprocess sandbox

Execution

Fly.io

Deploy

The clock's already running.

Pick a challenge, out-prompt the timer, and see how the judge scores you.

Enter the arena