name: RuiC-card-skill description: Generate interactive 3D holographic collectible-card websites from a user description or reference image, using layered artwork, Blender and Three.js. Includes project-local Blender installation, reusable parallax materials and browser verification. Host-agnostic — any harness that runs a multimodal model can use it, since the model draws the layer artwork and inspects the rendered frames. Use when the user asks for 全息闪卡, 镭射卡, 3D 卡牌网页, a holographic collectible card site, or an editable card.blend.

RuiC Card Skill

Turn the user's description or uploaded reference into a finished, editable Blender card and an interactive Three.js page. Preserve the requested subject, style, typography and destination. This skill contains code and text only; generated artwork belongs in the user's output project. It is host-agnostic — nothing in it is specific to one harness or one vendor's model — and it needs a multimodal model: you produce the layer artwork and you look at the rendered frames, so the host has to be able to generate and inspect images.

The layer stack

All layers share one canvas, normally 1024×1536 portrait. Each layer has one job and one depth, and the depth differences are what make the card read as stacked planes instead of one flat picture.

Give the layers clearly different depths. Two layers at the same depth fuse into one plane and the parallax stops reading as depth; that is the single most common way a card ends up looking flat.

Working sequence

1. Establish the output project and a compact card specification: subject, title, subtitle, technique or tagline, edition, palette, medium, destination. Infer harmless missing details and state them. Read references/art-direction.md for the per-layer quality bars and reference handling. Do not assume every card is anime, Japanese, or a copy of the example that inspired this skill.
2. Produce the layers. For an uploaded reference, inspect it first and preserve the requested identity and composition; use the available image tool and never silently substitute an API that needs a key or a paid service. Subject and text (and the optional effects overlay) must be genuinely transparent RGBA; the background must be fully opaque. Save them under the output project's assets/ as subject.png, background.png, lineart.png, text.png, plus effects.png when used. Keep every layer on the same canvas. Never package generated artwork into this skill.
3. Derive lineart.png from the same source as the subject — the same draw calls, or an edge/threshold pass over the finished subject — so the contours register exactly. Never ask for a fresh line drawing from a text description: a regenerated pose drifts and produces glowing outlines in the wrong place.
4. Write <project>/card-config.json from references/config.example.json: metadata, the asset paths, and one depth per layer. Then run scripts/validate_assets.py <project> and inspect the result by eye as well as numerically — equal canvases, real transparency where it is required, dark contours on white for the line art, and no empty layers. A painted checkerboard is not transparency: validate_assets.py converts a painted board to a true alpha channel through scripts/checkerboard_to_alpha.py (logged in asset-validation.json), but asking the tool for a genuine RGBA PNG is still the better outcome.
5. Run scripts/run_pipeline.py --project <project>. It validates, locates Blender and installs an official portable copy into <project>/tools/ when absent, builds the editable scene, renders the preview, exports the geometry, copies the viewer from assets/web-template/, and installs the viewer's dependencies. Python with Pillow, Node.js with npm, and network access for the first Blender download are required. Read the errors instead of retrying the same blocker.
6. Start the site with node <project>/web/server.mjs, keep it alive, and open the URL it prints. Read references/verification.md, then run node scripts/verify_web.mjs <project>: it launches its own headless Chromium and its own copy of the viewer, drives drag, flip, reset, zoom, the keyboard, every slider (including 特效景深), the finish options, the screenshot download, a ~390 px viewport and the reduced-motion setting, compares real captured frames, writes <project>/verification/report.json plus screenshots, and exits non-zero on any failure. It needs no GUI and runs as root, so a container or CI session is not a reason to skip it. A green report is evidence about controls, not about looks: still open the page and look at the render, both tilts and the layers yourself. A ready flag proves none of that.
7. Deliver the working URL, the source project, the editable .blend and the renders. Say plainly that the Blender node graph is rebuilt in GLSL for the browser — glTF carries geometry and material roles only — so the two are close but not pixel-identical. Publish a website or repository only when that is asked for. For a requested skill ZIP, run scripts/package_skill.py: it enforces a text-only allowlist, and output projects, artwork, models with embedded images, credentials, dependencies and caches never belong in it.

Non-obvious invariants

• All image planes are created in XY with object rotation X=90 degrees. Keep that rotation unapplied; faking it in the mesh breaks the local axes the parallax maths depends on.
• The shared group is 视差效果 with inputs 缩放 and 深度 and one output. Defaults: subject scale 1.25 and subject depth 0.4, background depth −0.25, text scale 1 and depth 0. A fixed layout safety mapping (safeArea) can hold room for typography; keep it separate from the user's parallax controls.
• UV centring plus a normal transform is not enough for view-dependent parallax. Transform the viewing direction into the card plane, divide by a bounded normal component, and apply it as a signed depth offset — the same expression on both sides: (p − 0.5)·s + 0.5 + uView.xy / max(|uView.z|, 0.35) · d · 0.14.
• The face mixes the background and subject BSDF by the subject's alpha. Keep metallic=1 and roughness=1 where the design asks for it, keep the subject BSDF emission black, and never bury a bad material under extreme emission.
• Foil is a material, not a painting: mapped bands (wave scale ≈0.55, distortion 7, mapping Y ≈32°) with their own pattern-image mapping, Multiply/Add, a pink→yellow→blue→white ramp and Overlay. The spectrum phase has to follow the viewing angle — a phase driven only by time looks like a looping video, not like foil.
• Stripe emission and the desaturated, thresholded line emission stay separate nodes. The line node may use strength 40 but must be sparsely masked, or it eats the printed detail underneath. Stars combine Voronoi distance-to-edge with animated noise. Card sides get their own material slot, and the compositor glow stays high quality.
• Do not bake the holographic effect into the artwork. If the picture already contains rainbow sweeps and glitter, the material has nothing left to do and the parallax planes turn into soup.
• Export real card geometry from the scene — never a flat screenshot presented as an imported model. The browser contract is the material names web_front, web_edge, web_back, web_gold (relief mode adds web_subject, web_effects, web_text), and the page composites the image layers with UV formulas that match the Blender graph.
• glTF's Y-up conversion changes Blender's local axes. Compute uView in the canonical card root frame rather than the converted front mesh's frame, and restore the exported V coordinate exactly once. Test both turn directions: one direction will smear the UVs or invert the parallax if the frame is wrong.
• The viewer exposes window.__holo (ready, config, renderer, root, uniforms, reset, flip, getState) for testing, and falls back to a CSS-3D card when WebGL is unavailable. Use the hook for scripted checks; a page that renders does not prove the shaders compiled.

Resources

• scripts/run_pipeline.py: one command from prepared assets to a served card — validate, Blender build, render, GLB export, viewer assembly, dependency install.
• scripts/ensure_blender.py: official-release discovery, SHA-256 validation, project-local extraction. If the release index is unreachable but a verified package already sits in <project>/tools/, extracting it there is enough — the script returns as soon as it finds tools/blender*/blender.exe.
• scripts/build_card.py, scripts/export_web.py: the editable Blender scene (parallax groups, foil, ink glints, stars, glow) and the geometry/material-role export.
• scripts/validate_assets.py, scripts/generate_typography.py: layer validation (including the optional effects layer) and accurate transparent typography.
• scripts/verify_web.mjs: end-to-end browser verification over the DevTools Protocol — locates a Chromium-family browser (an explicit --browser/RUIC_BROWSER wins, then the usual install locations, PATH, and Playwright's own downloads; on Linux it passes --no-sandbox --disable-dev-shm-usage so root and container runs work), starts the viewer on a free port, exercises every control in a desktop and a 390 px pass, compares captured frames, and writes <project>/verification/report.json plus screenshots. Node only, no dependencies.
• scripts/checkerboard_to_alpha.py, scripts/test_checkerboard.py: deterministic conversion of checkerboard-painted fake transparency, with regression tests.
• assets/web-template/: the responsive Three.js viewer: drag, wheel zoom, flip to the back, auto-play, four finishes, and sliders for 镭射 and the layer depths (画面比例 / 画面景深 / 特效景深 / 底纹景深). It ships as a single bundled file so per-module URLs cannot be blocked; rebuild it with assets/web-template/bundle.sh (bun, or esbuild when bun is absent) after editing app.js.
• scripts/package_skill.py: text-only allowlist, content checks and ZIP verification for sharing the skill itself.

✦ RuiC Card Skill

一个通用 Agent Skill：不挑宿主，多模态模型都能用——把小时候文具店门口那种会闪的全息卡复刻到浏览器里。 说一句话，就得到一张会随视角流光、带层次景深的 3D 闪卡网页，外加一个可以随便改的 Blender 工程。

小时候买不起的闪卡，现在你想印谁就印谁。这是你的私人卡牌工坊。

它是一台全自动闪卡生产线：你给一句话或一张参考图，剩下全交给模型——

1. 画四层图：主体、背景、线稿、文字，画在同一块画布、同一套坐标里
2. 搭 3D 场景：Blender 建卡牌几何，把四层图按景深在空间里"撑开"——主体前凸、背景后缩
3. 铺全息材质：镭射彩虹的相位跟着视角走，转到哪闪到哪；烫金 / 银箔 / 珠光 / 原画四种质感
4. 组装网页：Three.js 查看器打包成单文件，起本地服务，打开页面实测拖拽、翻面、滑块、手机布局
5. 交付：网页链接 + card.blend 源工程 + 四层透明 PNG + card-config.json 配置 + 渲染图

技术栈就三样：Blender（官方便携版自动安装，不碰系统环境）+ Three.js（按同一套 UV 公式重建材质）+ Python 流水线（画完图之后一步到位）。skill 本体只有代码和文字，装上就能用；你生成的画作、工程、模型全部待在你自己的项目目录里。

🎬 演示

调整前 · 图层分散

初版层距：主体、特效、文字隔得较开，转动时明显"散开"：

▶ 观看调整前完整视频

演示效果由 DeepSeek V4.1 Flash 实现。

调整后 · 层距收紧（现默认）

调整后：各层收近一档，卡片整体更紧凑，仍保留层次景深——这是当前出厂默认效果：

▶ 观看调整后完整视频

演示效果由 DeepSeek V4.1 Flash 实现。

✨ 特性

• 一句话出卡：描述或参考图 → 分层图 → 配置 → 流水线 → 网页，全程自动，你只负责想
• 不挑宿主、不挑模型：不绑定某一家——流水线里模型要自己画四层图、自己打开渲染帧判断对不对，所以只要是多模态模型（能出图、能看图），配任何读 SKILL.md 的宿主都能跑通
• 真 3D 层次景深：主体前凸、背景后缩，层与层随视角错开，不是一张平面贴图
• 视点流光：镭射彩虹的相位跟着视角走，转到哪闪到哪；烫金 / 银箔 / 珠光 / 原画四种卡面质感
• 浏览器里随便玩：拖拽旋转、翻面、景深/光泽/画面比例滑块，手机横竖屏都适配
• 出厂自带验收：node scripts/verify_web.mjs <项目> 自己拉起无头浏览器和本地服务，把拖拽、翻面、缩放、键盘、五个滑块、四种质感、截图下载、390px 窄屏、减动效逐项跑完，并且比对真实画面帧（不是只看滑杆读数变没变），报告和截图落进 verification/。macOS / Windows / Linux 都能跑，无 GUI 的容器与 root 环境同样直接可用
• 免装 Blender：官方便携版自动下载、SHA-256 校验后装进项目目录，不污染系统环境
• 网页零依赖请求：查看器打包成单文件（three + 图标全部内联），广告拦截插件无从下手；即使浏览器关了硬件加速，也有 CSS-3D 分层兜底，绝不会白屏
• 可编辑交付：card.blend 真工程 + 透明分层 PNG + card-config.json，想改哪层改哪层
• 纯文本 skill：本体只有代码和文字，一键打包 ZIP 交付，无二进制、无凭据、无缓存

🚀 快速开始

安装

skill 本体就是 SKILL.md + Markdown + 纯 Python/Node 脚本，不绑定任何宿主，也不绑定任何一家模型——只要宿主能读 SKILL.md、并且跑的是多模态模型（能生成图像、能看图），把它放进宿主的 skills 目录就能用（目录名即 skill 名；通用约定是 ~/.agents/skills/，别的宿主换成自己的 skills 目录即可）。

环境

• 模型：多模态（要能出图、也要能看图判断渲染帧；纯文本模型做不了这两步）
• Python 3.9+ + Pillow（ensure_blender.py 用了 Path.is_relative_to，3.6 / 3.8 上会直接报错）
• Node.js + npm
• Blender 不用自己装——流水线自动把官方便携版放到 <project>/tools/，校验 SHA-256。官方源在你的网络里被墙时，用 RUIC_BLENDER_BASE 指向镜像即可（例：RUIC_BLENDER_BASE=https://mirrors.aliyun.com/blender/Blender4.5/）；校验值取自同一个源，所以镜像要自己信得过
• 验收脚本需要一个 Chromium 内核浏览器：系统装的 Chrome / Edge / Chromium 都行，Playwright 缓存里的 Chromium 也认，还可以用 RUIC_BROWSER 或 --browser 指定路径。无 GUI 的 Linux（含 root / 容器）可以直接跑：脚本在 Linux 上自动补 --no-sandbox --disable-dev-shm-usage

开口

"用 RuiC-card-skill 给我做一张水墨风的锦鲤闪卡，文字用书法体，编号 No.001"

或者上传参考图：

"照这张图做一张闪卡，保留人物和构图，背景换成星空"

它会先把卡片规格说给你听，然后开工：画四层图 → 生成文字层 → 写配置 → 跑流水线 → 起本地服务 → 打开页面实测拖拽、翻面、滑块和手机布局 → 交付。

你会收到

🎬 可以拿它做什么

• 猫主子的传说卡：上传照片加一句"传说稀有度、金边框"，拖一拖，猫往前凸、背景往后退
• 独立游戏卡组：一个角色一句描述，战士、法师、盗贼、Boss 批量出货，每张卡一个配置
• 团队纪念卡：头像当主体、部门色当背景、Slogan 当文字层，网页链接一发大家翻一下午
• 节日仪式感：背面写一句话，对方翻到背面的那一刻，闪光效果拉满
• 发布会彩蛋：产品卡一个链接"扫码看会闪的那种"，观众当场转起来
• 材质实验场：card.blend 里镭射、星光、线稿发光都是独立可调节点，想学怎么"闪"就打开它

⚙️ 工作原理

一句话进去，一张会闪的卡出来，中间是一条全自动流水线：

flowchart LR
    A[一句话 / 参考图] -->|"分层提示词<br>art-direction"| B[四层图<br>subject·background·lineart·text]
    B --> C[card-config.json<br>卡片规格]
    C --> D[run_pipeline.py<br>一键流水线]
    D -->|"validate_assets<br>透明/对齐体检"| E{体检通过?}
    E -->|"否，重新生成"| B
    E -->|是| F[Blender 便携版<br>自动安装 + SHA-256 校验]
    F --> G[card.blend<br>可编辑视差场景]
    G --> H[card.glb<br>几何 + 材质名契约]
    H --> I[web-template<br>Three.js 查看器单文件打包]
    I --> J[本地网页 4173<br>拖转/翻面/滑块/手机布局]

核心链路要点：

• 四层图共用一套 UV 公式：主体、背景、线稿、文字在 Blender 和网页里按同一公式合成，所见即所得
• 视差不是简单贴图：把视角方向变换进卡面坐标系、除以有界法向分量，再按带符号景深偏移 UV，才有真正的"层与层错开"
• 全息镭射相位跟着视角走：转到哪闪到哪，而不是只随时间循环
• glTF 搬不动节点图：Blender 的自定义材质图没法经 glTF 直传，网页端用同一套公式重建 shader，并打包成单文件——多个小模块请求会被广告拦截插件误伤，单文件无懈可击
• Blender 装进项目里：官方便携版按 SHA-256 校验后解压到 <project>/tools/，项目自带环境、互不干扰

🔧 可以调的旋钮

流水线出厂就是一套顺手的参数，也都留了口子：

• 视差强度：主体默认 scale 1.25 / depth 0.4，背景 depth -0.25，想更"跳"就往上加
• 特效层：assets/effects.png 可选，独立景深（effectsDepth，网页里就是「特效景深」滑块），叠在人物之上、文字之下，适合花瓣/火星/藤刺这类装饰
• 卡框与文字：都放 text.png。网页对文字层不做视差，所以边框会稳稳钉在卡边
• 镭射条纹：条纹密度、扭曲度、角度，以及粉-黄-蓝-白的渐变
• 线稿发光：强度和遮罩密度，从"淡淡勾边"到"霓虹描边"
• 星光：Voronoi 尺度 + 动画噪声，从零星几颗到满天星
• Blender 界面语言：默认简体中文，存在项目本地配置里，一句话可换

📁 目录一览

RuiC-card-skill/
├── SKILL.md                    # 宿主读的"操作手册"
├── references/
│   ├── art-direction.md        # 分层画图的提示词写法、参考图处理
│   ├── config.example.json     # 卡片配置示例
│   └── verification.md         # 交付前的验收清单
├── scripts/
│   ├── ensure_blender.py       # 自动获取官方 Blender 便携版
│   ├── build_card.py           # 生成可编辑的 Blender 场景
│   ├── export_web.py           # 导出卡片几何
│   ├── generate_typography.py  # 精确的透明文字层
│   ├── validate_assets.py      # 四层图体检（棋盘格假透明图自动转真 alpha）
│   ├── checkerboard_to_alpha.py # 棋盘格底确定性抠透明（附回归测试）
│   ├── run_pipeline.py         # 一键流水线
│   └── package_skill.py        # 纯文本打包成可分享的 ZIP
└── assets/
    └── web-template/           # 响应式 Three.js 查看器

skill 本体只有代码和文字，轻得很。你生成的画作、.blend、模型都待在你自己的输出项目里。

📦 打包分享

想发给朋友或放进仓库：

python scripts/package_skill.py RuiC-card-skill --out ~/Desktop/RuiC-card-skill.zip

按白名单只打包文本文件，打出来的 ZIP 干干净净，拿走就能用。

赞赏支持