AGENT SKILL

3d-object

A product or object in 3D — built from shapes, extruded from a logo, or loaded from your own .glb — lit like a studio shot, turning on a turntable, ready to export as a still, a transparent cut-out or a turntable video.

License 许可
MIT
最近更新
2026-11-14
Stars
★ 1.1k
Forks
104

01 前言 Front Matter

od
{ "mode": "prototype", "craft": { "requires": [ "color", "anti-ai-slop" ] }, "preview": { "type": "html", "entry": "index.html" }, "platform": "desktop", "scenario": "marketing", "design_system": { "requires": false }, "example_prompt": "Make a 3D hero shot of our speaker, a soft cylinder with a glowing ring on top, in our brand colors, turning slowly — then export a transparent PNG and a turntable MP4." }
name
3d-object
tags
[ "3d", "three.js", "webgl", "product shot", "turntable", "glb", "gltf", "logo", "3D 模型", "产品图" ]
emoji
🧊
en_name
3D Object
zh_name
3D 物体
category
3d
scenario
marketing
triggers
[ "3d", "3d object", "3d model", "product shot", "turntable", "glb", "gltf", "3d logo", "hero render", "3D 模型", "产品渲染" ]
description
A product or object in 3D — built from shapes, extruded from a logo, or loaded from your own .glb — lit like a studio shot, turning on a turntable, ready to export as a still, a transparent cut-out or a turntable video.
en_description
A product or object in 3D — built from shapes, extruded from a logo, or loaded from your own .glb — lit like a studio shot, turning on a turntable, ready to export as a still, a transparent cut-out or a turntable video.
zh_description
3D 产品或物体:用基本形体搭建、由 Logo 挤出, 或载入你自己的 .glb 模型;摄影棚光照与转台, 可导出静帧、透明抠图或转台视频

03 SKILL.md 正文

3D Object

Intent. One object, shown well: a product, a logo or a simple form, lit like a studio photo, on a quiet stage, turning slowly. The page is a single self-contained three.js scene. People use it as a hero visual, export a still or a transparent cut-out for slides and pages, or a turntable clip for social.

This is not a 3D editor, a game scene or text-to-3D generation. If the user wants a model of something that doesn't exist yet and can't be built from simple shapes, say so and offer a procedural stand-in.

1. Pick the object path

  • Procedural: build the object from three.js geometry: LatheGeometry for anything turned (bottles, speakers, cups, lamps), ExtrudeGeometry for flat shapes with depth, boxes and cylinders with bevels (RoundedBoxGeometry from the addons) for devices. Combine a few parts in a Group. Aim for the silhouette and proportions first; small details (a groove, a light ring, a seam) make it read as a product.
  • Extruded logo: parse the brand's SVG with SVGLoader, turn each filled path into shapes (SVGLoader.createShapes), and extrude with a bevel. Flip y (SVG's y points down), center the geometry, and scale it to a fixed size. Strokes aren't extruded; ask for a filled version of the mark if that's all there is.
  • The user's model (.glb / .gltf): load it with GLTFLoader. Copy the file under assets/ next to the design (with a shell copy: it's binary) and load assets/<name>.glb; if you can't copy it, use its path relative to the design. Set up DRACOLoader with the pinned decoder path (below) so Draco-compressed models work. Never assume the model's units, origin or orientation: measure its bounding box, rest it on the ground, center it, and frame the camera from its size (the GLB example's stage() does exactly this). If the file is over 50 MB, tell the user it will load slowly and suggest npx @gltf-transform/cli optimize in.glb out.glb --compress draco first.

2. Build the scene

Use exactly these pinned URLs, through an import map:

<script type="importmap">
{ "imports": {
  "three": "https://cdn.jsdelivr.net/npm/three@0.160.0/build/three.module.js",
  "three/addons/": "https://cdn.jsdelivr.net/npm/three@0.160.0/examples/jsm/"
} }
</script>

Draco decoder path: https://cdn.jsdelivr.net/npm/three@0.160.0/examples/jsm/libs/draco/gltf/.

Don't write three.js script tags from memory: build/three.min.js no longer exists (removed in r160), nor does examples/js/… (removed in r148), so <script src=".../three.min.js"> 404s, THREE is undefined and the page shows nothing. Use the import map above and import * as THREE from 'three' in a <script type="module">, with addons from three/addons/….

  • Canvas: one full-bleed <canvas id="stage" data-od-webgl> passed to WebGLRenderer({ canvas, antialias: true, preserveDrawingBuffer: true }). Size it from the canvas's CSS box on load and on resize. On tall, narrow canvases (aspect below 1), frame the object again so it fills the portrait frame as it fills the desktop one: widen the fov and move the camera back only as far as the object needs (the fitView snippet in the 3d-scene skill does this from the object's bounding sphere). The check measures how much of each frame the object fills.
  • Color: outputColorSpace = SRGBColorSpace, toneMapping = ACESFilmicToneMapping.
  • Materials: physically based (MeshStandardMaterial / MeshPhysicalMaterial), with colors read from the design system's tokens (getComputedStyle(document.documentElement).getPropertyValue('--accent')) so the object matches the page: the body from --surface or --fg, one detail from --accent. Choose roughness and metalness for the real material (matte plastic 0.6–0.9 rough, brushed metal metalness 1 at 0.25–0.4 rough, glossy paint with clearcoat). Keep a user model's own materials unless asked to recolor it.
  • Light: the studio setup by default: studio light from a generated room (scene.environment = new PMREMGenerator(renderer).fromScene(new RoomEnvironment(), 0.04).texture, envMapIntensity about 0.6), a shadow-casting key DirectionalLight #fff4e6 at intensity 3 from 45° to the side and 40° up (soft shadows: PCFSoftShadowMap, shadow.radius about 6), a fill #e6eeff at 0.9 from the other side and a rim at 2 from behind, exposure 1.0. When the brief sets a scene (daylight, a lamp, a fire, dusk), use that setup from the 3d-scene skill's light table instead; a light you can see is the key and points at the object. The object is the brightest thing in the frame that isn't a light. When real reflections matter (glass, chrome, glossy paint) or the brief names a setting's light (a sunset, a studio, a forest), find an HDRI with find_open_design_3d_assets (type: "hdri", free CC0 files from Poly Haven) and light from it: new RGBELoader().loadAsync(url) from three/addons/loaders/RGBELoader.js, then scene.environment = pmrem.fromEquirectangular(tex).texture. Load it inside ready. The same tool finds PBR textures (wood, metal, fabric) and models when shapes alone won't do. List every remote file you use in the head, so a standalone export (embedAssets: true) or a Claude artifact can carry it inside the page: <script type="application/od-assets+json">[{ "source": "polyhaven", "id": "studio_small_09", "url": "https://dl.polyhaven.org/…/studio_small_09_1k.hdr", "license": "CC0" }]</script>.
  • Glass: don't use transmission: it renders unreliably in software WebGL, where checks and exports run, and costs an extra render per frame. Use a thin clear shell instead (MeshPhysicalMaterial with transparent: true, opacity 0.1–0.2, roughness about 0.03, clearcoat: 1, envMapIntensity about 1.6, depthWrite: false), and light it from an HDRI so it shows real highlights.
  • Stage: a large ground plane with ShadowMaterial (opacity 0.12–0.2) under the object. When the brief asks for a setting (a hall, a desk, a shelf), build it as an environment: put it in its own group with userData.odStage = true (the object itself is never inside that group), and keep it darker, softer and less detailed than the object so it never competes with it. Review views frame the object and leave the stage out. scene.background is the --bg color, or null for a transparent cut-out.
  • Controls: OrbitControls with damping, no pan for single objects, maxPolarAngle just under 90° so the camera never goes below the floor, and min/max distance around the framed distance.
  • Budget: at most about 200k triangles and 2048 px textures, so it stays smooth in software rendering. Lathe and torus segments of 64–160 are plenty.
  • Transparent option: a TRANSPARENT constant at the top. When true: alpha: true on the renderer, scene.background = null, and a class on <html> that makes the page background transparent and hides any caption.
  • Page: the scene fills the window. A short caption (product name, one line) in the design system's type is fine; keep it in a corner and out of the object's way.

3. The ready contract (required)

Exports, the visual check and motion export rely on it:

const start = performance.now();
window.odScene = { rotate: true, rotationSeconds: ROTATION_SECONDS, duration: ROTATION_SECONDS, ready: undefined, scene, camera, renderer, controls };
function pose(now) {
  const turn = window.odScene.rotate ? (((now - start) / 1000) / ROTATION_SECONDS) * Math.PI * 2 : 0;
  product.rotation.y = AUTHORED_ANGLE + turn;
}
window.odScene.ready = (async () => {
  await loadEverything();            // the model, textures, fonts for 3D text
  pose(performance.now());
  renderer.render(scene, camera);    // a direct render, not inside requestAnimationFrame
})();
renderer.setAnimationLoop((now) => { pose(now); controls.update(); renderer.render(scene, camera); });
  • ready resolves only after every asset has loaded and one frame has been drawn with a direct render() call. Motion export runs the page on a virtual clock where animation frames only advance on its schedule, so a ready that waits for an animation frame would never resolve there.
  • The turntable angle is computed from time (performance.now() or the loop's timestamp), never incremented per frame, so a rotation takes exactly ROTATION_SECONDS at any frame rate.
  • rotate = false returns the object to AUTHORED_ANGLE. Still exports set it, so choose the authored angle as the best three-quarter view.
  • scene, camera and renderer are required: the check uses them to render the review views (overview, close-up, low angle, wireframe) and to read the renderer's triangle and draw-call counts. Pass controls too when there are orbit controls. Set focus to the object (its group): the check measures the still's framing and light on it at desktop and mobile. Optional: subject (the group to frame in review views, when the stage flag isn't enough), a [x, y, z] focus instead when only the close-up should aim at a detail, and views ({ low: { position: [x, y, z], target: [x, y, z] } } to replace a computed view).

4. Register and check

  1. register_open_design_artifact with kind html and this skill as the source.
  2. check_open_design_artifact. Exports and the check switch on software WebGL for pages with data-od-webgl or a three.js import, so this works without a GPU. Fix every webgl finding: an error means no WebGL context (a script error before the renderer was made, or a canvas without the attribute's renderer); a warning means the canvas is a flat color (the model didn't load: see broken-asset; the camera is inside or away from the object; or ready resolved before anything was drawn). Fix a webgl-budget warning by instancing or merging repeated meshes and lowering segment counts. Fix every framing warning (the object too small, cut off, or a flat object edge-on in the still) and light warning (the object darker than its surroundings) before the critique.
  3. The second image is the review sheet: overview, close-up, low angle and wireframe in one picture. The text adds pixel metrics and renderer counts. A clean check only means the page renders; it says nothing about whether the object looks right. That's the critique's job.

5. Critique (required before you say it's done)

A design that passes the check can still be plainly wrong: a board with no squares, glass that reads as grey plastic, a model floating above its stand. After the check has no errors, get a critique from someone who didn't build it:

  • Who critiques. In Claude Code, launch a subagent with the Agent tool. In VS Code, use the runSubagent tool when it's available. Otherwise (Codex, or no subagent tool), do a separate critic pass yourself: judge only the review sheet and the brief, not what you meant the code to do.
  • The critic's brief. Give it the user's brief, the entry path and the scorecard below. It must call check_open_design_artifact itself and judge from the images. It returns:
    • a score for each category;
    • any automatic failure;
    • at least three flaws, each naming the view it's visible in, where in the frame, and a concrete fix (e.g. "low view: the board floats above a smaller hexagonal base; make the base wider than the board, or put the board on legs");
    • after finding a flaw, a check for others of the same kind (if one piece floats, look at every piece).
  • Loop. Fix the flaws, check again, and get a new critique. Stop when the scorecard passes, or after three critique rounds. Then tell the user the final scores and any flaws that remain. Don't paste the critique itself unless they ask.

Scorecard. Score each category 0 (placeholder), 1 (basic), 2 (finished) or 3 (showcase):

Category 1 2 3
Form primitives that only suggest the object authored silhouette and true proportions, recognizable as a dark shape memorable silhouette with functional detail where the camera looks
Materials flat colors, one roughness everywhere each surface reads as its real material (glass, wood, metal, stone) rich, cohesive materials with wear, trim or texture
Light even light, washed out or murky intentional key, fill and contact shadow; the form reads cinematic light that still shows the form clearly
Composition object small, cut off or off-center without reason object fills the frame well at the authored angle and on mobile framing that adds drama or story

Judge composition on the desktop and mobile screenshots, which show your own camera. The review views are framed by the check itself, so they always fill the frame. | Stage | missing, or competes with the object | quiet ground or environment that supports the object | an environment that adds context and depth without stealing focus | | Fidelity | misses or misreads what the brief asked for | everything the brief names is present and recognizable | details the brief implied but didn't spell out | | Technical | check errors, or budget warnings ignored | clean check, within budget, loops seamlessly | clean, light and fast to export |

Automatic failures (the scorecard fails whatever the scores):

  • something the brief names is missing or unrecognizable;
  • the object is an unrefined primitive placeholder;
  • fog, darkness or glow stands in for missing geometry;
  • the object floats above, or passes through, what it rests on (low and wireframe views);
  • the object fills less than 40% of the desktop screenshot's height without a reason the brief gives;
  • the check reports an error.

Pass: every category at least 2, and no automatic failure.

6. Export

  • Still: export_open_design_artifact with format: "png" (or a preset). Add selector: "#stage" to capture only the scene.
  • Transparent cut-out: set TRANSPARENT = true in the file (or make a copy that does), then export with format: "png", transparent: true. A warning that no pixel is transparent means the page or scene still paints a background.
  • Turntable: format: "mp4" (or "gif" for a small loop, "webm"). With odScene.duration set to ROTATION_SECONDS, the export is exactly one turn and the loop is seamless, with no duration to pass. Keep ROTATION_SECONDS between 6 and 12; a GIF loops best at 6–8 seconds and 480–720 px wide.
  • The scene loads three.js from the CDN, so viewers need a connection, including for standalone exports.

04 同仓库更多 Skills

feimacode
feimacode/open-design-agent-kit

youtube-clipper

YouTube clip generation and editing with automated workflows — pull source video, slice highlights, add captions, and export.

★ 1.1k
feimacode
feimacode/open-design-agent-kit

writing-guidelines

Review docs/prose for Writing Guidelines compliance. Use when asked to "review my docs", "check writing style", "audit prose", "review docs voice and tone", or "check this page against the writing handbook".

★ 1.1k
feimacode
feimacode/open-design-agent-kit

wpds

WordPress Design System. Apply WordPress's official design tokens, typography, and component patterns to themes and sites.

★ 1.1k
feimacode
feimacode/open-design-agent-kit

weread-year-in-review-video-template

WeRead-inspired HyperFrames video template for vertical annual reading reports, personal reading dashboards, book-note recaps, and shareable year-in-review stories. Use when users want a 9:16 HTML-to-MP4 reading report with warm paper texture, editorial Chinese typography, book-page metaphors, data highlights, and deterministic motion.

★ 1.1k
feimacode
feimacode/open-design-agent-kit

web-design-guidelines

Review UI code for Web Interface Guidelines compliance by the Vercel engineering team. Covers layout, typography, color, motion, and accessibility for product UI.

★ 1.1k
feimacode
feimacode/open-design-agent-kit

web-clone

网站复刻 / 克隆方法论。USE WHEN 用户说 复刻网站、克隆网站、clone website、抄个站、仿站、 照着这个站做一个、reproduce site、还原某个网页效果、把这个站搬下来改成我的、 复刻某个交互/WebGL/Canvas/Three.js 效果。提供「先拿真源码 → 判路径 → 逆向拆解 → 搭工程 → 替换内容」的可移植决策树,覆盖静态站 / React-Vue-Next 内容站 / WebGL-Canvas 重前端站三大分支,并强制核对任何 AI 二手分析里的可执行代码。

★ 1.1k