Leonxlnx / athenaion

# Athenaion
A massive ancient Greek world built entirely in code with three.js. There are no external assets: every mesh, texture, shader, tree, statue and waterfall is generated procedurally. Claude Code (Opus 5.5) built it over about 2 days to match 9 reference images, and it was rendered in 4K on an integrated laptop GPU (AMD Radeon 780M).
## What's in it
- An acropolis on a limestone cliff massif with terraces, retaining walls, a podium and a grand stair
- A Parthenon-style Doric temple with gilded acroteria, sculpted pediments, metopes and owl banners
- Colossal Athena statues and dozens of smaller statues, all sculpted as signed distance fields and meshed
- 57 traced waterfalls with spray, mist and foam, plus a lagoon, a river with rapids and a sea with planar reflections
- Aqueducts and bridges, a city of procedural houses, a harbour with ships, a tholos on an island
- 149k cypress, olive, pine and broadleaf trees and shrubs, with ivy, grass and flowers
- An eroded snowy mountain range, volumetric clouds, cloud banks and aerial perspective
- A giant fallen marble head and a ruined portico in the foreground
- An offline renderer with jittered accumulation (AA, soft shadows, motion blur), AgX tone mapping, local tone mapping, bloom and colour grading
| | |
|---|---|
| | |
| | |
## Numbers
| | |
|---|---|
| Total time | ~45 h wall clock (Oct 16, 03:36 to Oct 18, 00:34) |
| Building and iterating | ~27 h across 9 multi-agent rounds |
| Agents | 55 subagents (up to 12 in parallel), ~147 agent-hours |
| Final rendering | 4K stills plus a 54 s 4K 30fps video, ~13 h of GPU rendering |
| Code | ~45k lines of JavaScript in 168 modules |
| Tokens | ~11M output tokens, ~3.8B cache-read tokens, ~161M cache-write tokens |
| API-equivalent cost | ~$1.8k to $2.3k at Opus 5.5 API prices (actually run on a Claude Max subscription) |
## References
The 9 AI-generated reference images the world was built to match are in `refs/`. Left: reference, right: final 4K render.
## How it was made
The whole session prompt is in docs/PROMPT.md.
1. Claude Code wrote the engine first: the render harness, atmosphere, sky, shadows and the accumulation pipeline.
2. It wrote an art bible and module contract (`docs/`), then ran 12 specialist agents in parallel: terrain, water, sky, architecture, acropolis, aqueducts, statues, ruins, trees, flora, city and props. Each agent rendered its work, compared it against the references and fixed it in a loop.
3. Further rounds of director, look-dev and module agents tightened composition, colour, detail density and waterfalls, using side-by-side comparisons and colour statistics against every reference.
4. The final video was rendered frame by frame (8 or 4 samples per frame with motion blur), cut with crossfades and scored.
## Running it
 [code] 
Then open `http://localhost:8080/?shot=1&quality=preview`. Keys 1-9 switch shots and the mouse orbits.
Heavy geometry (terrain meshes, statues, the giant head) is not committed. The first load