With Apologies To Taito

Score

Drop in

← → aim · space fire · R restart

What this is

old-project-watt was a Windows-only C++ game built on fixed-function OpenGL, GLEW, DevIL and a hand-rolled ASE model loader — 5,622 lines that last built against a Visual Studio project file. The bundled GLEW 1.7 dates it to 2011 or so; it reached git in 2019 under the commit message “old work from god knows how long ago”. It never ran anywhere you could link to.

This is that game rewritten as a three.js app: a renderer-agnostic engine with unit tests, a WebGL front end, and a static build that deploys anywhere. The window title is the original’s, and so is the geometry.

Carried over from the C++

The playfield numbers were lifted straight out of gameRules.h rather than re-derived:

Ceiling height117.3
Grid8 × 13
Max arm angle85°
Half-width44.0
Rot. sensitivity0.35

The original also declared clusterSize = 4 and never used it. This version pops at three, like the arcade game.

The bubbles

Thick glass in three.js is harder than it looks: turning up ior and thickness on a MeshPhysicalMaterial gets you real refraction, but it smears whatever is suspended inside into mush. So the casing wall is faked as an alpha-tested radial band that lives in the transmission buffer — opaque enough to be refracted by the shell, while the shell itself stays at a low IOR so the Cognition logomark inside stays legible.

Each logo hangs from a pivot at the bubble’s centre. A landing or a pop sends out a shockwave: the impulse decays with distance and arrives later the further out it goes, then a damped spring rights it again.

Controls

Aim← → or mouse
FireSpace or click
RestartR

How this got made

One session, start to finish: eleven prompts from Jetha, about half an hour of wall clock, four commits. No code was written by hand on the human side — the loop was a prompt, a screenshot, and a correction.

Prompts11
Wall clock~30 min
Commits4
Unit tests23
C++ retired5,622 lines

“rewrite it as a three.js puzzle bobble”

Read the C++ for the rules, moved it to legacy/, scaffolded Vite + TypeScript + three.js, and wrote the engine — hex grid with per-row parity, flood-fill pops, orphan drops, ceiling pressure — behind 23 unit tests before any of it was drawn.
“deploy it to watt.jethachan.net”

Kept the build a plain static dist/ with Node pinned, so a Pages project needs no config beyond npm run build. A second agent got the domain and the credentials; this one never saw them.
“glass-walled spheres with a Cognition logo inside”

Physical glass, an environment map for something to reflect, and the official logomark rasterised to a canvas texture. First attempt put the logo on a transparent quad, which promptly disappeared: transmissive materials sample an opaque snapshot of the scene, so anything transparent inside them is invisible. Alpha-testing the quad put it back.
“two-thirds the size, and rock it on impact”

Logo pivot at the sphere centre, an impulse that falls off exponentially with distance from the landing, and a delay of distance ÷ wave speed so the ripple visibly travels outwards.
“the glass doesn’t look very thick”

Correct. Sweeping IOR and thickness through the live page showed the tradeoff plainly: every setting thick enough to read as glass turned the logo to soup. Fixed by faking the wall as geometry in the transmission buffer instead of asking the material for it.
“we’ll need a game complete and game over modal”

Replaced the text banner with proper end-of-round cards — score, bubbles left, a focused Play again button, and R still bound.
“the sphere is clipping through the launcher”

It was: the muzzle was narrower than a bubble. Widened it past the bubble radius and reseated the loaded shot in the cradle.
“make the rocking more obvious”

Tilting a near-symmetric mark barely changes its silhouette, so the motion was there in the numbers and invisible on screen. Adding a twist about the view axis — and a softer, longer spring — made the ripple readable.
“the framerate here is balls”

On the agent’s VM, yes: no GPU, just software rendering at about 3 fps. Motion was verified numerically instead — peak tilt in radians, decay over time — with stills for the look.
“add a little sidebar”

This.

What prompting was like

Short and visual. Nothing in the eleven prompts specified an API, a material, or a file — they described what the game should look and feel like, and the corrections were the kind you’d give an artist: thicker, more obvious, it’s clipping. The work in between was mostly measuring: rendering candidate settings in the live page, screenshotting them side by side, and picking.