An experiment in how far coding agents can actually go.
Not a serious game — an academic exercise. The interesting artefact is not the game; it's that a small boy with an enormous head, the cave he's trapped in, the music, and the story he's told were all produced in a conversation, from 27 prompts, with no code typed by hand.
The game underneath is real and playable: you are alone in a cave holding the only light, and the passages out are sealed behind rock that looks exactly like every other wall. Your torch is the instrument that finds them.
npm install && npm run dev # http://localhost:5173or, with nothing installed but Docker:
docker compose up --build # http://localhost:8080| Prompts from the human | 27 |
| Tool calls by the agent | 793 |
| Elapsed | ~25 hours over two days |
| Lines of code typed by hand | 0 |
The prompts are almost entirely art direction and bug reports rather than instructions — "the head is not moving", "it sounds sort of like rushing air", "the tin whistle piece is weird", "lol maybe not go down again". The whole game is specified in about two pages of plain English.
The agent drove Blender over MCP to author the meshes and rig, Playwright
to drive and verify the real game in a real browser (286 calls), and a great deal
of bash to measure things that cannot be eyeballed.
That last part is the finding. Every serious bug in this project looked fine:
- The character was "animated" for a whole session while 1750 of its 1792 vertices had no vertex group — the clips played, the bones moved, and nothing on screen budged.
- The choir "sounded ethereal" while measuring 1.48× noise to pitched signal. It was wind, not a voice.
- Foot-slide was "a bit off" at 5.91×.
None of those were found by looking. They were found by measuring, and the project now runs on that rule.
📖 How this was built — which tools did which work, when, why, and the four lessons that cost the most.
📜 The full transcript — all 27 prompts and everything that followed each one, including the wrong turns, the misdiagnoses, and the corrections. This is the actual point of the repo.
The torch is the instrument. The flame stirs and cools toward blue-white as a hidden way comes into range; the wall dissolves, the passage opens, and something ancient comes up out of the floor and into him.
The way up has to be earned. A cavern keeps its staircase until every passage hidden in it is open — and not by locking it: until then the way up is simply absent, unlit and off the map.
The ending takes the controls away. He walks the last steps into the daylight and comes apart upward, and Claude writes the last of his story knowing everything that happened in the run.
Clips are recorded at a slightly closer camera than the game ships with, so the boy reads at GIF size. Nothing else is altered — same lighting, same rules, driven through the real keyboard input.
Which is really the second half of the experiment — how much of a game can be generated rather than made?
- The caves are built by cellular automata in the browser at load and re-rolled until they pass their own design rules. There is no level file.
- The geometry is built from those grids as three.js
BufferGeometry; the 2.4MB cavern mesh is gone entirely. - The music has no audio files. The score is synthesised, and the choir that rises as you near something worth finding is a physically-modelled vocal tract — a Kelly–Lochbaum waveguide in an AudioWorklet.
- The story is written by Claude while you play, from a log of where you actually went and what you actually found. Never the same twice.
- The art is generated by Python in Blender. No mesh was modelled by hand.
Three caverns, climbed rather than descended, because the end-game is escape. Built with three.js. Runs in any modern browser; no credentials or Blender install needed to play.
Everything technical is in docs/, laid out so a coding agent can
navigate it — AGENTS.md is a table of contents rather than a
manual, and CLAUDE.md is symlinked to it.
| docs/HOW-THIS-WAS-BUILT.md | building it with coding agents |
| docs/transcript/ | the full session transcript |
| docs/ARCHITECTURE.md | how the system fits together, and the rules that hold it |
| docs/SETUP.md | toolchain, optional API credentials, Blender |
| docs/WORKFLOWS.md | how to make a change and verify it |
| docs/DEPLOYMENT.md | the container |
| docs/systems/ | story, audio, rendering, assets |
| docs/adr/ | why the load-bearing decisions were made |
MIT — do what you like with it. Every asset in this repo is generated
by the scripts in blender/; no third-party art is vendored.
Style reference: Kenney kits — chunky proportions,
flat facets, muted palette. No Kenney assets are used; every mesh here is
generated by the scripts in blender/. The loader is generic, so dropping real
Kenney props in is straightforward.



