Imported from gabrielmoreira/agent-skills-mirror (mirrors/repos/Kayforkind@skill-slice/skills/three-js-web-game/SKILL.md). Install upstream with npx skills add gabrielmoreira/agent-skills-mirror --skill three-js-web-game. Copyright stays with the author.

Three.js Web Game — high-fidelity browser games, zero install

Build games that run by opening a URL. The stack is Vite + TypeScript + three (npm). The output of npm run build is a static folder you can host anywhere — GitHub Pages, Netlify, Cloudflare Pages, itch.io HTML5, or a local npx serve. No engine editor, no export templates, no installer. The browser talks to the GPU through WebGL, so the RTX 3090 Ti is used automatically — no CUDA/driver work needed; Chrome is both the dev runtime and the verification target.

Priority order for this skill: visuals first, then feel, then features. A flat-lit, un-graded Three.js scene looks like a tech demo from 2012. The difference between that and "best of the best" is almost entirely: PBR materials + environment lighting + tone mapping + postprocessing + art direction. That path is the Procedure below — do not skip it.

Related skills: creative/comfyui (generate the art this game consumes), svg-quality-audit / svg-asset-qa (gate SVG assets before ingest), godot-2d-game-development (the native-engine alternative — use that when the target is desktop builds, not web).

When to Use

• Building a browser-delivered game with Three.js (no install, no engine editor).
• Converting AI-generated art or SVG vault assets into a playable WebGL scene.
• Standing up a Vite + TypeScript + three project from scratch.
• Achieving high visual fidelity: PBR materials, IBL, tone mapping, postprocessing.
• Designing game architecture: state machines, systems, input-as-intent.
• Performance tuning a Three.js game for 60 fps in Chrome.
• Verifying a Three.js game visually via browser CDP and screenshot analysis.

Do NOT use this skill when: the target is a native desktop/mobile binary (use godot-2d-game-development), or when you need a full-featured game engine editor with scene graphs and visual scripting.

Prerequisites

• Node.js installed (for npm create vite).
• Chrome installed — it is both the dev runtime and the verification target.
• GPU: The browser uses WebGL automatically; an RTX 3090 Ti or similar is utilized without driver/CUDA work.
• Windows host (primary): Commands below assume PowerShell. Use npm directly; no WSL required.
• Related skill outputs (upstream):
  • creative/comfyui — generates AI art (albedo, reference images) this game consumes.
  • svg-quality-audit / svg-asset-qa — run on vault SVGs before ingest. Broken viewBoxes and invisible-fill bugs waste a render cycle to discover in-engine.
• Reference files: If this skill ships a references/ directory, load those files when the Procedure calls for deep dives into specific subsystems (e.g., postprocessing pass configs, KTX2 compression details). If a scripts/ directory exists, use those helper scripts for asset conversion tasks as noted in the steps below.

Procedure

Step 1 — Scaffold the project (Vite + TypeScript + three)

Why Vite: instant HMR (edit a material parameter, see it without reload), TypeScript out of the box, an asset pipeline that hashes/bundles textures and audio, and a production build that is plain static files. Hand-rolling <script src="three.min.js"> costs you all of that and breaks the moment you need addons (loaders, postprocessing live in three/addons/, which needs a bundler or import maps).

npm create vite@latest ink-and-spirit -- --template vanilla-ts
cd ink-and-spirit
npm i three howler
npm i -D @types/three @types/howler
npm run dev        # -> http://localhost:5173, HMR on
npm run build      # -> dist/ (static, deploy anywhere)
npm run preview    # serve dist/ locally to sanity-check the production build

Project layout — systems, not one giant file:

index.html                  # one <canvas> host, nothing else
public/
  assets/
    textures/               # AI-art albedo/normal/roughness maps (PNG/WebP/KTX2)
    hdri/                   # .hdr environment maps
    sprites/                # rasterized SVGs, UI atlases
    audio/                  # ogg/mp3
src/
  main.ts                   # bootstrap only: create Game, start loop
  game/
    Game.ts                 # owns renderer/scene/camera/composer, ticks systems
    states/
      StateMachine.ts       # generic FSM
      FreezeState.ts  FlightState.ts  FightState.ts  FawnState.ts
  systems/
    render.ts               # renderer + composer setup (Step 3 lives here)
    input.ts                # pointer/keyboard -> intent events
    loader.ts               # LoadingManager, texture/audio helpers
    audio.ts                # Howler wrapper, ducking, mute
  scenes/                   # per-chapter scene builders

public/ files are served at / verbatim (use for HDRIs, audio, large textures); anything imported from src/ gets hashed and bundled. Both work — public/ is simpler for game assets.

Step 2 — Minimal render loop (src/main.ts)

import * as THREE from 'three';

const renderer = new THREE.WebGLRenderer({ antialias: true, powerPreference: 'high-performance' });
renderer.setPixelRatio(Math.min(devicePixelRatio, 2));   // cap DPR — 4K + DPR 2 melts fill rate
renderer.setSize(innerWidth, innerHeight);
renderer.toneMapping = THREE.ACESFilmicToneMapping;      // filmic response — see Step 3
renderer.toneMappingExposure = 1.0;
document.body.appendChild(renderer.domElement);

const scene = new THREE.Scene();
const camera = new THREE.PerspectiveCamera(50, innerWidth / innerHeight, 0.1, 100);
camera.position.set(0, 1.5, 4);

const clock = new THREE.Clock();
function tick() {
  requestAnimationFrame(tick);
  const dt = Math.min(clock.getDelta(), 0.05);  // clamp: tab-switch returns a huge delta
  update(dt);                                   // game logic — dt-scaled, never frame-counted
  renderer.render(scene, camera);               // becomes composer.render() in Step 3
}
tick();

addEventListener('resize', () => {
  camera.aspect = innerWidth / innerHeight;
  camera.updateProjectionMatrix();
  renderer.setSize(innerWidth, innerHeight);
});

Step 3 — The visual fidelity path (do this BEFORE adding gameplay)

This is the checklist that takes a scene from "gray boxes" to matching AI-generated reference art. Apply in order; each step is visible on its own.

3.1 Tone mapping + color space (the foundation)

ACESFilmicToneMapping (set in Step 2) gives filmic highlight rolloff instead of clipped whites. Since three r152 the renderer outputs sRGB by default — your job is tagging inputs correctly (Step 3.3). Get this wrong and everything is washed-out or crushed and no amount of light tweaking fixes it.

3.2 Environment lighting (IBL) — the single biggest visual win

PBR materials look dead under point lights alone. Give the scene an environment map and every MeshStandardMaterial/MeshPhysicalMaterial picks up believable ambient light and reflections:

import { RGBELoader } from 'three/addons/loaders/RGBELoader.js';

new RGBELoader().load('/assets/hdri/moody_forest_2k.hdr', (tex) => {
  tex.mapping = THREE.EquirectangularReflectionMapping;
  scene.environment = tex;                    // lights ALL PBR materials
  scene.environmentIntensity = 0.9;           // grade to taste
  // scene.background = tex;                  // optional: also show it
});

Free HDRIs: polyhaven.com (2k is plenty for lighting; use 4k+ only if it's the visible skybox).

No HDRI that fits the mood? Neutral studio IBL with zero assets:

import { RoomEnvironment } from 'three/addons/environments/RoomEnvironment.js';
const pmrem = new THREE.PMREMGenerator(renderer);
scene.environment = pmrem.fromScene(new RoomEnvironment()).texture;

Then add one key light (DirectionalLight, with shadows) for direction and drama. Resist adding five lights — one key + IBL + postprocessing beats a light-salad every time.

3.3 PBR materials from AI art

Generate/derive three maps per surface (ComfyUI can produce albedo directly; normal/roughness can be derived with tools like Materialize or a normal-from-height node):

const tl = new THREE.TextureLoader(manager);

const albedo = tl.load('/assets/textures/ink_paper_albedo.png');
albedo.colorSpace = THREE.SRGBColorSpace;               // COLOR data -> sRGB. Mandatory.

const normal = tl.load('/assets/textures/ink_paper_normal.png');   // DATA -> leave linear
const rough  = tl.load('/assets/textures/ink_paper_rough.png');    // DATA -> leave linear

const mat = new THREE.MeshPhysicalMaterial({
  map: albedo, normalMap: normal, roughnessMap: rough,
  roughness: 1.0,        // multiplies the map
  sheen: 0.3,            // nice for cloth/paper — fits an ink & paper aesthetic
});

HARD RULE: albedo/emissive = sRGB; normal/roughness/metalness/AO = linear (default). The #1 Three.js visual bug is a missing or extra colorSpace = SRGBColorSpace.

3.4 Postprocessing (bloom, AO, grade) — where "game" becomes "art"

import { EffectComposer } from 'three/addons/postprocessing/EffectComposer.js';
import { RenderPass } from 'three/addons/postprocessing/RenderPass.js';
import { UnrealBloomPass } from 'three/addons/postprocessing/UnrealBloomPass.js';
import { OutputPass } from 'three/addons/postprocessing/OutputPass.js';

const composer = new EffectComposer(renderer);
composer.addPass(new RenderPass(scene, camera));
composer.addPass(new UnrealBloomPass(
  new THREE.Vector2(innerWidth, innerHeight),
  0.35,   // strength — subtle; > 0.6 looks like 2008
  0.8,    // radius
  0.85,   // threshold — only bright things bloom
));
composer.addPass(new OutputPass());   // applies tone mapping + sRGB — MUST be the last pass

// in tick(): composer.render() instead of renderer.render()
// in resize: composer.setSize(innerWidth, innerHeight)

• Bloom: emissive ink glow, spirit effects. Threshold high, strength low.
• AO: SSAOPass from addons works; the n8ao npm package (N8AOPass, drop-in) looks better and runs faster — worth it for a visuals-first game.
• Color grade: cheapest big win is a LUT — grade a screenshot in any photo editor, export a LUT, apply with LUTPass + LUTCubeLoader. This is how you lock the game to the reference art's palette without touching materials.
• Also consider ShaderPass with a vignette, and film grain for the ink aesthetic.

3.5 Match the reference art — the actual workflow

1. Pick 2–3 AI-generated reference images (the target look).
2. Build the scene with the steps above, screenshot it (see Verification).
3. Put screenshot and reference side by side (or vision_analyze both) and diff: palette, contrast, light direction, texture density, edge quality.
4. Adjust in this order: environment/exposure → key light → material maps → grade/LUT → bloom.
5. Repeat. HMR makes each iteration seconds, not minutes.

For 2.5D — a strong option given a large 2D asset library — put AI art on lit planes (MeshStandardMaterial + alpha), add parallax layers, fog (scene.fog), and particles. Full-3D fidelity is not required to look stunning; Hollow Knight is planes.

Step 4 — Ingest assets from the SVG vault and AI art

Gate first: run svg-quality-audit on any vault SVGs before ingest — broken viewBoxes and invisible-fill bugs waste a render cycle to discover in-engine.

SVGs as sprites/UI — rasterize, don't parse (default path). Browsers rasterize SVG natively:

async function svgToTexture(url: string, size = 512): Promise<THREE.Texture> {
  const img = new Image();
  img.src = url;
  await img.decode();
  const c = Object.assign(document.createElement('canvas'), { width: size, height: size });
  c.getContext('2d')!.drawImage(img, 0, 0, size, size);
  const tex = new THREE.CanvasTexture(c);
  tex.colorSpace = THREE.SRGBColorSpace;
  tex.anisotropy = 8;
  return tex;
}
// use on THREE.Sprite (billboard) or a PlaneGeometry + MeshBasicMaterial({ transparent: true })

Caveat: the SVG root needs width/height or a viewBox for Image to size it — the vault audit checks this. Rasterize at the largest on-screen size you'll need (power of two).

SVGs as geometry (special cases): SVGLoader from three/addons parses paths into ShapeGeometry/ExtrudeGeometry — real 3D extruded icons/glyphs. Use sparingly; complex SVGs triangulate badly.

Pure HUD/menus: plain HTML/CSS overlaid on the canvas beats in-scene UI — SVGs drop in as <img>, and it's accessible and crisp for free.

AI art (PNG/WebP): load via TextureLoader, tag sRGB. For big scenes, compress to KTX2:

toktx --genmipmap --t2 --encode uastc out.ktx2 in.png

Load with KTX2Loader — stays compressed in VRAM, 4–8× memory saving. If a scripts/ helper exists for batch KTX2 conversion, use it here.

3D models: glTF only (GLTFLoader + DRACOLoader) — never OBJ/FBX on the web.

Step 5 — Game architecture

• One Game class owns renderer, scene, camera, composer, and an array of systems with update(dt). main.ts constructs it and starts the loop. Nothing else touches the renderer.
• Input as intent: input.ts translates pointer/keyboard into semantic events (approach, withdraw, hold, soothe) — states consume intents, never raw keycodes. This also makes remapping and touch support trivial later.
• State machine — the 4 Fs are literally states:

interface GameState {
  enter(from: string | null): void;
  update(dt: number): void;
  exit(): void;
}

class StateMachine {
  private states = new Map<string, GameState>();
  private current: GameState | null = null;
  private name: string | null = null;

  add(name: string, s: GameState) { this.states.set(name, s); }
  transition(to: string) {
    this.current?.exit();
    const prev = this.name;
    this.name = to;
    this.current = this.states.get(to)!;
    this.current.enter(prev);
  }
  update(dt: number) { this.current?.update(dt); }
}

Each F-state owns its camera behavior, palette/grade, audio bed, and available intents — FreezeState.enter() might desaturate via the LUT pass, slow the camera, and muffle audio (Howler low-pass). Transitioning states is the emotional storytelling.

• Camera: fixed/authored cameras (lerped between anchor points) fit a narrative game far better than free orbit. Keep OrbitControls behind a ?debug URL flag for development only.
• Loading: one THREE.LoadingManager for everything; onProgress drives a styled loading screen (in the game's aesthetic — first impression matters); start the loop only onLoad.
• Audio: Howler for music/SFX (handles autoplay-policy unlock on first gesture); THREE.PositionalAudio only if you need true 3D positioning.

Step 6 — Trauma-informed design → mechanics

• Recognition over reliving: abstract/symbolic representation (ink, weather, spirit forms), never literal reenactment. The visuals-first approach serves this directly.
• The 4 Fs are protective, not failures: every state has valid moves and forward paths. No fail states, no game-overs, no forced repetition of distressing beats. "Losing" redirects, never punishes.
• Player choice as the growth arc: state transitions are chosen by the player (intent events), never forced by timers or scripted takeovers. Agency is the mechanic.
• Pacing and exits: pause always available, instant and judgment-free; autosave at every transition; a persistent "breathe"/grounding interaction that is always valid input.
• Soft feedback: grade shifts, audio warmth, light — not klaxons, screen shake, or red flash.

Step 7 — Performance (target: 60 fps in Chrome)

The 3090 Ti is not the bound — the browser's single JS thread and fill rate are. Budget: < 300 draw calls, < 500k triangles, DPR capped at 2.

• Read renderer.info.render.calls / .triangles every frame in a debug overlay from day one.
• InstancedMesh for anything repeated (particles, foliage, ink drops): 1000 objects → 1 draw call.
• Texture atlases for sprites/UI (many small textures → many draw calls); share materials so Three.js can batch state changes.
• LOD (THREE.LOD) only if you build large 3D environments; irrelevant for 2.5D.
• KTX2 textures (Step 4) for VRAM; frustumCulled stays on (default).
• Profile with Chrome DevTools Performance tab (long frames → is it JS or GPU?) and Spector.js extension (per-draw-call GPU capture).
• Postprocessing costs one full-screen pass each — bloom + AO + grade + output is fine; ten passes is not.

Pitfalls

Verification

1. npm run dev (background), confirm the terminal reports the localhost URL.
2. Open in Chrome via browser_cdp → navigate to http://localhost:5173.
3. Check the console first (browser_cdp console logs): Three.js warns loudly — texture size, shader errors, missing assets all land here before they're visible.
4. Screenshot via browser_cdp.
5. vision_analyze the screenshot against the acceptance bar: correct palette vs reference art, no black/magenta surfaces (missing texture), no washed-out haze (color space bug), UI legible.
6. Before shipping: npm run build && npm run preview and repeat steps 2–5 against the preview URL — dev and prod builds can differ (asset paths, base URL).

Asset-level QA belongs upstream: svg-quality-audit before SVGs enter public/assets.

Quick reference

# scaffold
npm create vite@latest <game> -- --template vanilla-ts
cd <game> && npm i three howler && npm i -D @types/three @types/howler

# develop / verify / ship
npm run dev                      # localhost:5173, HMR
npm run build && npm run preview # production sanity check
npx serve dist                   # or deploy dist/ to any static host / itch.io HTML5

Visual pipeline order: ACES tone mapping → HDRI environment → one key light → PBR maps (sRGB albedo, linear data) → EffectComposer (Render → Bloom → AO → LUT → OutputPass last) — then iterate against reference art with screenshot + vision diff until they match.