phaser-asset-pipeline

name	phaser-asset-pipeline
description	setting up or changing the…

Phaser 4 Asset Pipeline

Where the raw art comes from (read this first)

This skill packs raw art; it does not conjure it. The inputs under assets/src/** are real, licensed source art you sourced first — they do not magically pre-exist, and they are not procedural generateTexture rectangles. Before you can pack anything, use [[phaser-asset-sourcing]]: it covers where game art comes from (curated CC0 packs like Kenney and Ninja Adventure → deriving CC0 art → asking the human), the strict CC0-or-equivalent license gate with assets/LICENSES.md evidence, provenance via a committed ingest script, and why generateTexture placeholders are tracked art debt rather than a finished state. The rule of thumb: source → assets/src → pack (this skill) → typed keys.

Overview

The pipeline turns committed source art into runtime form — it is a build step, not a substitute for having art. Raw source art lives in assets/src/** (PNGs, audio, font sources); a build step packs it into public/assets/** (atlases, audio sprites, bitmap fonts) and codegens typed key constants into src/assets.ts. The payoff: a missing or renamed asset is a compile error, not a green square or silent-missing audio at runtime. No raw string asset / scene / event keys anywhere — they all come from the generated module (the official loading-assets skill covers loading those keys at runtime).

Layout

Path	Role
assets/src/sprites/**	Raw individual PNGs — the art source of truth (sourced per [[phaser-asset-sourcing]])
assets/src/audio/**	Raw audio clips (wav/ogg sources)
assets/src/fonts/**	Font sources for BMFont generation
assets/LICENSES.md	License evidence per source pack (CC0 quote + evidence URL) — see [[phaser-asset-sourcing]]
scripts/ingest-assets.mjs	The provenance record — how each frame was carved from its source pack into assets/src (re-runnable)
public/assets/**	Generated packed output (atlases, audiosprites, fonts, pack.json) — git-ignored or committed as a build artifact
src/assets.ts	Generated typed key constants — never edited by hand
scripts/pack-assets.mjs	The pipeline driver (runs the three packers + codegen)

Step 1: texture atlases with free-tex-packer-core

free-tex-packer-core packs many PNGs into atlas pages plus a JSON manifest, in a Node script (no GUI, runs in CI):

// scripts/pack-assets.mjs (excerpt)
import { packAsync } from "free-tex-packer-core";
import { glob } from "glob";
import { readFile, writeFile } from "node:fs/promises";

const images = await Promise.all(
  (await glob("assets/src/sprites/**/*.png")).map(async p => ({
    path: p.replace("assets/src/sprites/", ""), contents: await readFile(p),
  })));

const files = await packAsync(images, {
  textureName: "game", width: 4096, height: 4096,
  packer: "MaxRectsPacker", packerMethod: "Smart", // see gotcha below
  padding: 2, extrude: 1,                            // 1px extrude kills bleeding at integer scales
  allowRotation: false, allowTrim: false,            // pixel-art-safe (see below)
  detectIdentical: ,
  : , : , : ,
});
 ( f  files)  (, f. ?? f.);

The frame names in the manifest become the atlas frame keys you reference as generated constants. Atlas everything that renders together — loose images break sprite batching (the official loading-assets skill). Full-screen backdrops and parallax layers are the exception: copy them verbatim into public/assets/images and load.image them — they would bloat an atlas. Even this exception keeps the no-raw-string promise: the codegen step emits one generated constant per copied image (the Img record below), and the loader derives the file path from the key in exactly one place — a call site never spells out a path string.

Packer gotcha (verified in the reference implementation): use packer: "MaxRectsPacker" with packerMethod: "Smart". The other packer methods produce degenerate single-row strip atlases that waste enormous texture space and can blow past the max texture size. For pixel art, set allowTrim: false and allowRotation: false: trimming makes animation frames different sizes so anchors swim between frames, and rotation makes Phaser frame-flipping non-trivial. Add extrude: 1 to stop edge bleeding when the game renders at integer scales. Keep inputs sorted and emit no timestamps so the packed output is deterministic and can be committed + git diff --exit-code-checked in CI.

Step 2: audio sprites with audiosprite

audiosprite concatenates short SFX into one file (in multiple codecs) plus a JSON map of { start, end } per clip — one request, one decode, instead of dozens:

// package.json script
"pack:audio": "audiosprite --output public/assets/audio/sfx --export ogg,m4a --format howler2 assets/src/audio/*.wav"

Provide at least two codecs (ogg + m4a) so every browser can play one. Load with this.load.audioSprite(Audio.Sfx, "audio/sfx.json") and play by sprite key.

Step 3: bitmap fonts with BMFont

BitmapText is the performant choice for high-churn text (scores, timers) — the official loading-assets skill / [[phaser-i18n]]. Generate the .fnt (XML/JSON) + PNG page from a font source with a BMFont tool (msdf-bmfont-xml or the bmfont CLI) in the same script:

"pack:font": "msdf-bmfont -o public/assets/fonts/ui --font-size 42 --texture-size 1024 1024 assets/src/fonts/ui.ttf"

Load with this.load.bitmapFont(Font.UI, "fonts/ui.png", "fonts/ui.fnt").

Step 4: codegen typed keys into src/assets.ts

After packing, walk the generated manifests and emit as const key maps. This is the step that turns runtime failures into compile failures:

// scripts/pack-assets.mjs (codegen excerpt)
const atlas = JSON.parse(await readFile("public/assets/atlases/game.json", "utf8"));
const frames = Object.keys(atlas.frames); // raw keys, e.g. "player/idle"
const sfx = Object.keys(JSON.parse(await readFile("public/assets/audio/sfx.json","utf8")).spritemap);
const images = (await readdir("public/assets/images", { recursive: true }))
  .filter(f => f.endsWith(".png")).sort(); // standalone backdrops/parallax layers

const out = `// AUTO-GENERATED by scripts/pack-assets.mjs — do not edit.
export const Tex = { GameAtlas: "game" } as const;
export const Frame = { ${frames.map(k => `${toIdent(k)}: "${k}"`).join(", ")} } as const;
export const Img = { ${images.map(p => `${toIdent(p)}: "img-${p.replace(/\.png$/u, "")}"`).join(", ")} } as const;
export const Audio = { Sfx: "sfx" } as const;
export const Sfx = {  } as const;
export const Font = { UI: "ui" } as const;
export const SceneKeys = { Boot:"Boot", Preloader:"Preloader", MainMenu:"MainMenu", Game:"Game" } as const;
export const GameEvent = { ScoreChanged:"score-changed", EnemyDied:"enemy-died" } as const;
`;
 (, out);

Scenes import Tex, Frame, Img, Sfx, Font, SceneKeys, GameEvent — never an inline string. Rename art, re-run the pipeline, and every now-stale reference fails bun run typecheck immediately. (Scene/event keys are kept in the same generated module so they share the no-raw-string discipline; edit them via the codegen template, not by hand.)

Animation keys are the one deliberately hand-authored layer. Codegen cannot know how frames group into cycles, so animation keys live in a hand-written as const record (e.g. AnimKeys / FxAnims) whose frame sequences are built from the generated Frame constants — and the contract test below proves every frame those definitions reference exists in the committed atlases. The typed-key guarantee is therefore: generated constants for textures, frames, images, audio, and fonts; hand-authored typed constants + contract-test coverage for animation keys. Nothing is a raw inline string either way.

Step 5: wire into the build

The pipeline runs before dev and before build so generated output is never stale:

// package.json
"scripts": {
  "assets": "node scripts/pack-assets.mjs",
  "predev":   "bun run assets",
  "prebuild": "bun run assets",
  "dev":   "vite",
  "build": "vite build"
}

For fast iteration, a --watch mode on the pack script re-packs changed source art. In CI the pack step runs once before vite build; the build then hashes the generated files for immutable caching ([[phaser-build-deploy]]).

Project conventions

• assets/src/** is the source of truth; public/assets/** and src/assets.ts are generated — treat them as build output (re-runnable, not hand-edited).

• One atlas per render group, one audiosprite per SFX set, BMFont for hot text.

• A contract test asserts every frame name the game derives by convention (e.g. battlerWalkFrame(ref, dir, n), FX-strip frame lists, per-speaker portrait keys) resolves to a real entry in the committed atlas JSON under public/assets. The other constants in src/assets.ts (SceneKeys, GameEvent, Font, Audio, standalone Img entries) are plain generated literals already enforced by TypeScript at every use site, so they don't need a runtime check — only names built at runtime from a pattern can go stale silently without one. This keeps that narrower guarantee honest: a renamed or missing frame fails a headless Vitest assertion instead of rendering a silent black square. The test reads the packed JSON directly — zero Phaser ([[phaser-testing]]):

  const frames = new Set(Object.keys(
    JSON.parse(readFileSync("public/assets/atlases/battlers.json", "utf8")).frames));
  for (const ref of BATTLER_REFS)
    for (const dir of DIRS) {
      expect(frames).toContain(battlerIdleFrame(ref, dir));
      expect(frames).toContain(battlerAttackFrame(ref, dir));
      for ( n = ; n < ; n++)
        (frames).((ref, dir, n));
    }
  
  
  
  

Verification

The pipeline is verified by deleting a source PNG and re-running bun run assets: the corresponding generated constant disappears and bun run typecheck fails at every use site (proving keys are compile-checked). Then boot the game and confirm the atlas/audiosprite/font load with no FILE_LOAD_ERROR and render/play correctly.