agentleFS
cth9191cth9191/site-clone/skills/shader-extract/SKILL.md

shader-extract

Skill109 starsChanged 3 months agoMITView on GitHub

Extract, reproduce, and locally replay a single WebGL / WebGPU / Canvas / shader / animated-3D visual effect from a web page — the GPU-rendered surface that DOM/CSS cloning is blind to. Use whenever a page has a WebGL/WebGPU/canvas shader background, animated 3D hero, generative gradient, particle field, or Unicorn Studio / Spline / Three.js effect that must be captured and rebuilt as a runnable local module. This is the GPU track of a site clone (driven by clone-site) but also stands alone. Trigger on "extract this shader", "clone the WebGL background", "port this canvas effect", "rebuild this animated hero", "reproduce this Three.js scene locally". NOT for ordinary DOM/CSS cloning (use dom-clone).

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 1 is distinctive, and 5 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 5

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Read vendor/web-shader-extractor/SKILL.md and its references/replay-policy.md — that is the operating manual for capture → baseline → …

    line 50
  2. Very specific:

    Run the vendored state flow from CAPTURE_MINIMUM_TRUTH onward, writing evidence and the baseline into the given output dir.

    line 54
  3. Very specific:

    Honest labels: Every implementation-critical fact is SOURCE, PARTIAL, or GUESS.

    line 66
  4. Very specific:

    Never quietly upgrade a BEHAVIOR_REBUILD (lookalike) to sound like a source replay.

    line 68
  5. Very specific:

    No compensation tuning: Don't nudge brightness, speed, offsets, noise scale, or color to mask missing pipeline evidence.

    line 70
  6. Distinctive:

    Cross-origin honesty: If the canvas is cross-origin/tainted and pixel or shader readback is blocked, say so and fall back to a captured …

    line 76

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Shader Extract

Capture one GPU-rendered surface and rebuild it as a verified, runnable local effect module. The governing principle — inherited from the vendored engine this skill wraps — is evidence before implementation: capture what the page's GPU pipeline actually does (shaders, uniforms, pass order, textures, timing) and match it, rather than eyeballing the visual and hand-tuning a lookalike. A lookalike drifts; a capture-matched baseline is faithful and honest about its gaps.

This skill = intake adapter + a vendored evidence-gated engine

The heavy machinery lives in vendor/web-shader-extractor/ — an evidence-gated state machine (surface lock → source/runtime trace → capture → verified baseline → projectize) by lixiaolin94 (MIT; see ATTRIBUTION.md). Do not reimplement it. This wrapper does two things the vendored engine assumes were done elsewhere:

  1. Intake — accept a pre-locked surface from the clone-site coordinator so you skip straight to capture instead of re-scouting the whole page.
  2. Capture backends — make the concrete, research-verified capture tools explicit (references/capture-backends.md): Spector.js for WebGL1/2 and WebGPU Inspector for WebGPU. These are what actually recover a frame; the vendored policy files describe the discipline, these describe the tools.

Router: how you were invoked

Driven by clone-site (common): you receive a locked surface — selector, bounding box, guessed driver (three.js / unicorn-studio / spline / regl / raw WebGL), and an output dir like output/<host>/effects/<surface-id>/. The surface is already attributed, so skip scouting. Go straight to:

  1. Read vendor/web-shader-extractor/SKILL.md and its references/replay-policy.md — that is the operating manual for capture → baseline → projectize.
  2. Pick the capture backend by surface type (references/capture-backends.md): WEBGL1/2 → Spector.js; WEBGPU → WebGPU Inspector.
  3. Run the vendored state flow from CAPTURE_MINIMUM_TRUTH onward, writing evidence and the baseline into the given output dir.
  4. Return a mount-contract module (see below) plus the honesty label.

Standalone (user hands you a URL): you must lock the surface first. Run the full vendored flow from INTAKE (read vendor/web-shader-extractor/SKILL.md). The clone-site script scripts/surface-map.js is a fast way to enumerate and classify canvases if you want a head start, but the vendored TARGET_LOCK gate is authoritative.

Non-negotiables (carried from the vendored engine)

  • Honest labels. Every implementation-critical fact is SOURCE, PARTIAL, or GUESS; unlabeled = GUESS. The final module carries a fidelity verdict. Never quietly upgrade a BEHAVIOR_REBUILD (lookalike) to sound like a source replay.
  • No compensation tuning. Don't nudge brightness, speed, offsets, noise scale, or color to mask missing pipeline evidence. A visible gap gets recorded in known-gaps.md, not fudged away.
  • Baseline before projectize. Get a faithful, verified capture-baseline running first; only then refactor it into an editable module. Don't overwrite a verified baseline for cleanup.
  • Cross-origin honesty. If the canvas is cross-origin/tainted and pixel or shader readback is blocked, say so and fall back to a captured poster frame — don't invent a shader you couldn't read.

Output contract (so clone-site can mount it)

Beyond the vendored output/ layout (capture-baseline, editable-project, qa-report, known-gaps), expose the effect through the mount lifecycle the coordinator composites against:

export function mount(canvasEl, opts) { /* start renderer + rAF loop */ return handle }
export function unmount(handle)        { /* cancel rAF, loseContext, free GPU */ }
export function resize(handle, w, h, dpr) { /* resize drawing buffer + viewport */ }

Also emit a poster frame (a captured still) as a fallback for reduced-motion and for the case where reconstruction only reached poster fidelity. Report the fidelity label so the coordinator carries it into the final clone report unchanged.

Where to read next

  • references/capture-backends.md — Spector.js (WebGL) and WebGPU Inspector (WebGPU): what each captures, how to drive it, what's reliable vs guesswork.
  • vendor/web-shader-extractor/SKILL.md — the full state machine + reference router (surface discovery, target lock, evidence policy, replay policy, QA, Three.js/TSL reconstruction, Unicorn Studio / shaders.com adapters).
  • ATTRIBUTION.md — upstream authorship and MIT license.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.