One HTML file. Yard, metro, paint physics, HUD, undo: everything arrives in a single response body and plays. No engine fetch, no asset manifest, no second request.

I cut the atmosphere pass after building it completely. Seven perception-layer features: contact shadows under every structure, drifting light motes, lamp halos, god rays, a pulsing sun, a dragon shadow sweeping the ground, a breathing idle camera. All functional. All deleted.

What I chose against: the version of the game that screenshots better. What it cost: the stand-in-the-plaza-at-dusk shot, and hours of real craft I was proud of.

I would cut it again. Every one of those effects asked the eye to look at the rendering instead of at the paint. The game is the mark you make. Anything that competes with the mark is subtraction disguised as polish, even when it is beautiful, even when it works. Craft you can delete on purpose is worth more than craft you cannot.

The mark is computed from hand speed and can state, tuned until a fast stroke reads as fast and a slow one reads as deliberate. The velocity curve, the spray range, and the blur behaviour are not published.

The metro is a timer with a platform, not a path network. The yard is a grid with a few rules about which faces accept paint. Neither list is published: the feel is the product, and the feel does not transfer by reading.

Primitive: cloth(r,i), order 8, period 7. One boolean array read through the hero field, the valve instrument, this diagram, and the selvedge on the index spec block.

Contrast: bone on void 15.61:1; ash on void 5.35:1; crimson on void 5.54:1. Captures: NO RECEIPT, three plates running as notation. What this page cut: the stack string, the tuning table, and every screenshot it could not produce honestly.

Cold start on a slow connection was several seconds of silence before the first track. A player that pauses before it plays is broken at hello. And the browser suspends audio the moment it decides you are done listening; the UI kept advancing while the sound cut out.

I chose single-file architecture over a modular build system. No bundler, no framework, no package manager. The entire application lives in one long file.

What it cost: no hot module replacement, no tree-shaking, no component isolation. Every refactor is manual, and nobody else could navigate the file but me.

What it bought: zero build time, zero dependency drift, and the ability to reason about the whole application in one scroll. When audio dropped mid-playback I knew exactly where to look, because everything lives in the same scope. I would choose it again at this scale without hesitating. The tooling I declined exists to serve teams and codebases I do not have. The bug that mattered most, browser audio suspension, was found and fixed precisely because there was nowhere for it to hide.

Every track was remastered for instant start before the first line of player code was written. The encoding targets are not published. A keepalive keeps the audio thread awake through tab switches and lock screens; the mechanism is not published. The ten EQ bands land where they land because of listening, not measurement; the frequencies and gains are not published.

Primitive: cloth(r,i), order 8, period 7, shared across every surface of this portal. Contrast: bone on void 15.61:1; ash 5.35:1; crimson 5.54:1. Captures: NO RECEIPT, three plates as notation. What this page cut: the stack string and the encoding table.

A handheld 3D scene on a phone GPU. Every light, every post pass, every billboard spends from the same frame, and the frame budget is not negotiable: the scene either holds its rate in a hand or it does not ship.

I cut the party lights and kept the strobe.

The original build had three orbiting point lights cycling purples and reds. They lit the scene and flattened it; every character got the same wash. Removing them dropped ambient illumination by nearly half and forced the strobe, skylight, and searchlight to carry the composition.

What it cost: the safe version, where everything is visible all the time, plus the hours spent tuning lights I would go on to delete. What it bought: characters now emerge from darkness instead of floating on a lit stage, and the crimson reads as an event.

I would do it again. Lighting is casting: everything you light is something you said yes to. Saying yes to everything is how scenes go flat.

A small custom post chain replaces the off-the-shelf stack. The passes, their order, and their thresholds are not published. Camera energy feeds the separation effect directly; the curve that maps one to the other is not published.

The beat is synthesized. A pulse with phase offsets stands in for listening, tuned until the illusion holds. The tuning is not published.

Primitive: cloth(r,i), order 8, period 7, shared across every surface of this portal. Contrast: bone on void 15.61:1; ash 5.35:1; crimson 5.54:1. Captures: NO RECEIPT, three plates as notation. What this page cut: the stack string, the pass order, the gyro thresholds.

A matte black watch on a black hero. The product and the ground share a value, so separation must come from light alone. And the traffic is mostly phones, where the showpiece cannot afford to run.

I cut the showpiece on mobile.

The 3D deconstruction, seven real photographs suspended in space and exploded apart by scroll, is the best moment on the page. It is desktop-only. On phones, visitors get a static grid, and the pinned Blueprint section was removed entirely because it trapped mobile scroll.

What it cost: the majority of traffic never sees the most impressive thing I built. That is a real price for a portfolio piece.

I would do it again. A deconstruction that stutters is worse than no deconstruction. Parity of feel beats parity of features: a static image that scrolls perfectly reads as more considered than 3D that janks. The watch is hostile and precise; the page cannot be less disciplined than the object.

Smooth scroll and the animation system share one frame loop, which is why the page does not fight itself. The synchronization constants are not published. Black-on-black separation comes from a small lighting rig and a grading pass tuned for this one object; the rig and the grade are not published.

Primitive: cloth(r,i), order 8, period 7, shared across every surface of this portal. Contrast: bone on void 15.61:1; ash 5.35:1; crimson 5.54:1. Captures: NO RECEIPT, three plates as notation. What this page cut: the stack string, the scroll constants, the lighting rig.

Scroll-native 3D on mid-tier phones. Immersion without an orbit control, without an enter-3D step, without a physics bill the frame cannot pay.

I abandoned the physics-driven approach.

The initial spec called for a zero-gravity simulation: floating bodies responding to audio frequency data. It sounded impressive on paper. In practice it produced jittery movement on mobile, drained battery, and added serious weight to the bundle before any visual payoff.

The replacement was a camera on a fixed path with scroll-driven interpolation: trading the impressive sentence for the better feel. What it cost: the marketable feature. Physics-driven reads better in a portfolio than scrolls-perfectly, and I gave up the better sentence.

I would do it again. The feature that sounds best in a case study is rarely the one that feels best in the hand. The path is the experience; the physics was the pitch.

Artwork palettes are derived per release with a clustering pass chosen because it keeps skin tones and print inks honest. The specific algorithm is not published. The camera follows a fixed path driven by scroll progress; the path and the interpolation are not published.

Primitive: cloth(r,i), order 8, period 7, shared across every surface of this portal. Contrast: bone on void 15.61:1; ash 5.35:1; crimson 5.54:1. Captures: NO RECEIPT, three plates as notation. What this page cut: the stack string, the palette algorithm, the path constants. Credits: the four releases above, every artist named.