55 Subagents, 0 Asset Files: Claude Code Built a Whole Ancient Greek World in 45 Hours
- Level Advanced
- Setup1–3 hrs (estimated)
- Evidence Strong evidence
- CostNot stated in source
- No-codeNo
What they did
A developer had Claude Code build a huge ancient Greek world in three.js to match 9 reference images, with no outside assets: every temple, statue, tree and waterfall is drawn by code. Over about 45 hours and 9 rounds it ran 55 helper agents (up to 12 at once). Each one rendered its part, compared it with the references and fixed it. The result is about 45,000 lines of JavaScript, 4K stills and a 54-second 4K video, and the full session prompt is in the repo.
How it went
A large, fully procedural 3D scene that matches your reference images, built module by module.
Worth knowing
- Big job: about 11M output tokens, roughly $1.8k to $2.3k at API prices (he ran it on a Claude Max subscription) plus about 13 hours of GPU rendering for the final 4K video. Expect to steer it over several rounds.
Prompt
Cleaned version rewritten from the source — not a direct quote.
Build a [scene, for example a medieval harbour town] in three.js with no external assets: every mesh, texture and shader must be generated in code. Match these reference images: [attach images]. First write the render harness and a short art bible that splits the scene into modules (terrain, water, sky, buildings, trees, props). Then run one subagent per module in parallel; each renders its part, compares it side by side with the references and fixes it in a loop. Ask me before any step that will take hours of rendering.
Recommended
Try it in Grok Bot
- Copy the prompt (button below).
- Open Grok Bot chat and paste it. Replace any [bracketed] placeholders.
- Connect the listed tools (as needed) as connectors when Grok Bot asks.
- If the task repeats, ask: "Make this a daily routine."
Alternatives: the same prompt works in Claude Code, Claude, ChatGPT Agent, OpenClaw.
Set it up
Stack: Claude Code (Opus 5.5, subagents) + three.js
- Read the original source for full context.
- Connect the tools your agent needs.
- Paste the prompt and replace [bracketed] placeholders.
- Run once manually and check the output before scheduling it.
Generic suggested steps, not from the source.
Evidence
Public repo (about 150 stars in 2 days) with a detailed numbers table (time, agents, lines of code, tokens, cost), reference-vs-render comparisons, the full session prompt and run instructions
Source: Web · 2026-10-17 · Verified by us: not yet · Report broken or outdated