From one prompt to photoreal Manhattan: Matt Shumer's five-step agent loop for GPT-6 Astra 3D worlds
mattshumer_ · x · 2026-09-08
Matt Shumer has used GPT-6 Astra to build a photoreal 3D Manhattan in Unreal Engine, a 1:1 browser-walkable replica of his childhood neighborhood with a playable zombie survival game, and a simulated civilization of autonomous agents — pulling over 15M views on X.
He published the full method (prompts included):
- No silver-bullet prompt; the loop is fixed: Reference → Assets → Assembly → Critique → Ship.
- Anchor on real references: force the model to ground on real-world imagery (or GPT-Image-2 generations), not imagination.
- Sub-agents: Astra dispatches dedicated sub-agents to sculpt assets headlessly in Blender, then assembles scenes in Three.js or Unreal Engine.
- Blind critic: an unyielding critic that never sees the render sits between output and reference, kicking flawed work back until it passes.
A practical recipe for developers aiming to reproduce high-quality AI-generated 3D worlds and games.
Related event: Matt Shumer Shares Full GPT-6 Astra 3D World-Building Workflow(2 posts)→
More from coding & agent
- Super Smash Bros. Melee fully decompiled after 6 years, LLMs and Astra finished the job — mohbibi_ · 2026-09-08
- ChatGPT update reportedly breaks git worktrees as agents message across them — msg · 2026-09-08
- Dev slams multi-agent PR workflow: why ping my agent again for minor changes — steipete · 2026-09-08
- One-Step Code Generation via Continuous Diffusion and Trajectory Distillation — pengzhangzhi · 2026-09-08
- Dev building a top-down game with AI — BLUECOW009 · 2026-09-08
- Shannon 3.0: Open-Source AI Pentester With 47K Stars Now Maps Source Code Before Exploiting Live Apps — testingcatalog · 2026-09-08