棱镜闪卡工坊（Prism Card Forge）

把用户的一句话描述或一张参考图，变成一张可编辑的 Blender 卡牌与一个可交互
的 Three.js 网页。忠实保留用户指定的主体、风格、文字排版与用途。本技能只含
文本与代码：生成的素材属于用户的输出项目，绝不写回技能目录。

品牌特征：棱镜（Prism）——一束白光经棱镜色散成七彩，正是全息镭射卡
"不同视角折射出流动虹彩"的视觉本源。每一张卡都是一次"光的锻造"：
分层绘制、Blender 立体构建、实时折射着色，让平面的画在手中"活"起来。

本技能参考 Holo Card Studio（MIT 开源项目，见 LICENSE.md），用非 Codex 的新
模型重新实现：出图由本助手内置图像工具完成，不再依赖 Codex 内置画图；图层
流水线额外处理了该工具“无真实 alpha、图面右下带水印”的特性（抠图、去水印、
画布居中）。

什么时候用 & 怎么用（先读这一节）

什么时候选这个技能： 只要你想要"一张会发光、能拖拽转动、带音效的 3D
收藏卡网页"——全息镭射卡、棱镜卡、盲盒抽卡、角色卡/纪念卡/应援卡、带随机
种子玩法的 collectible card——直接开口即可，无需记任何命令：

• 「给我做一张机械猫全息收藏卡，赛博朋克风，标题写 NEON CAT，稀有度 SSR」
• 「按这张参考图 @image 做一张会闪的 3D 收藏卡网页」
• 「随机抽一张全息卡给我，要带宠物和音效，种子用 8888」

也可以指名道姓：「用棱镜闪卡工坊做……」。交付物固定为一个本地预览网址
（可拖拽/翻面/调镭射/听声效）、一张渲染图、可编辑的 card.blend 工程和
一条可复现的种子链接。

首次使用会发生什么： 安装技能本身不装任何东西。首次运行自动准备环境：
Blender（约 380 MB）优先从清华 TUNA 镜像下载，失败自动回落官方源，
支持断点续传——中断/超时后重跑同一条命令即可续传，不必从头下载；
Three.js 依赖由流水线自动 npm install。装好的 Blender 复用于之后的每次
使用，不会重复下载。全程约需几分钟（视网速），其余步骤全自动。

最快路径（agent 会自动执行，列出仅供了解）：

python <skill>/scripts/randomize_card.py --project <project>  # 掷一张随机卡（可选）
python <skill>/scripts/ensure_blender.py <project>            # 先后台跑：装 Blender
python <skill>/scripts/run_pipeline.py  --project <project>   # 建卡+渲染+组装网页
node <project>/web/server.mjs                                 # 打开 http://127.0.0.1:4173

文档地图（按需查，无需通读）：

Runtime environment (auto vs. on-demand)

Installing this skill installs nothing by itself. On first use the agent
prepares the environment while running the sequence below:

• Blender (~380 MB) — automatic. ensure_blender.py first reuses any
  system Blender, otherwise downloads an official SHA-256-verified portable
  build into <project>/tools — Tsinghua TUNA mirror first with automatic
  fallback to download.blender.org, resumable .part download and 3 retry
  rounds; a timed-out run just needs the same command re-run. Start it in
  the background early (step 2).
• Node dependencies — mostly automatic. The pipeline itself runs
  npm install --ignore-scripts in <project>/web/ (three.js). For the
  browser acceptance step the agent must install playwright-core once
  (npm install playwright-core --no-audit --no-fund) and drive the system
  Edge/Chrome channel — no Chromium download.
• Python + Pillow — on-demand. The helper scripts import PIL; when that
  fails the agent creates an isolated venv and pip install pillow once
  (python -m venv <env> && <env>/pip install pillow). Node/npm and a Python
  3.10+ interpreter must exist on the host; the agent may install them if
  completely absent.

Never claim completion when a prerequisite is missing; prepare it, then retry.

Working sequence

1. Establish project and card spec. Create <project>/ with assets/,
   copy references/config.example.json to <project>/card-config.json and
   fill title / subtitle / technique / tagline / edition / collection /
   description from the user's request. Infer harmless missing details and
   state them. Read references/art-direction.md for layered-image guidance —
   including the 宠物/配饰/动作/镜头 section that injects variety into the
   artwork itself. Do not assume every card is anime or based on the example
   images. Then roll the card's identity:
   python <skill>/scripts/randomize_card.py --project <project>
   (optionally --seed, or pin traits like --pet mecha-cat --rarity SR) —
   it writes a seed into card-config.json so every card is unique yet
   reproducible; quote the printed seed in the delivery. Users who ask for an
   exact look can skip this and set seed/variant by hand (see
   references/randomness.md).
2. Reserve the heavy downloads first. Start Blender install in the
   background: python <skill>/scripts/ensure_blender.py <project>. It places
   an official SHA-256-verified portable Blender under <project>/tools
   (~380 MB on Windows; download runs while artwork is generated). Inspect
   errors; the bundled script already carries the User-Agent fix for
   download.blender.org 403s.
3. Generate the artwork layers with WorkBuddy's image tool
   (text-to-image, 1024×1536; image-to-image for line art):
   - background → assets/raw/background.png (opaque; calm top/bottom 15%).
   - subject → assets/raw/subject.png (on a plain near-white backdrop).
   - Then run python <skill>/scripts/prepare_layers.py --project <project>
   to produce the real-alpha assets/subject.png, the watermark-patched
   assets/background.png, and the white-flattened assets/raw/subject_flat.png.
   Inspect the cutout visually and numerically; a painted checkerboard or an
   RGB-only file is NOT transparency.
   - line art → image-to-image from assets/raw/subject_flat.png with
   high input fidelity and an explicit "same pose, position, proportions,
   do not move" instruction; save to assets/lineart.png.
   - text → python <skill>/scripts/generate_typography.py <project>
   (real fonts on transparent PNG; Windows defaults to KaiTi, override via
   config.font).
4. Write metadata and validate. Confirm card-config.json; run
   python <skill>/scripts/validate_assets.py <project>; visually inspect
   transparency, registration of the line art against the subject, and that
   the typography zones are unobstructed. Regenerate drifting line art before
   continuing.
5. Run the pipeline. python <skill>/scripts/run_pipeline.py --project <project>
   — builds the editable Blender scene (card.blend), renders
   renders/hero.png, exports web/assets/card.glb, copies the Three.js
   template into <project>/web/ and runs npm install there. Python with
   Pillow plus Node/npm must be available. Do not claim completion or repeat a
   failing blocker; inspect the log instead.
6. Serve and verify in a real browser. Keep alive:
   node <project>/web/server.mjs (port 4173). Follow
   references/verification.md; run scripts/verify_card.mjs,
   scripts/verify_workshop.mjs and scripts/verify_audio.mjs
   (playwright-core with the system Edge/Chrome channel, no Chromium download)
   and prove with pixel differences that dragging, flipping, the foil slider,
   重掷, camera presets, pet and accessory cycling actually change the render —
   that the same ?seed= reproduces an identical card AND soundprint — and
   that the sonic engine unlocks on gesture, follows the card back and really
   schedules voices. Test around 390px width without overflow. A ready flag
   alone does not prove shaders compiled.
7. Deliver. Working URL (server kept alive), source project, editable
   card.blend, renders/hero.png, plus the card's seed link
   (…?seed=<n>，复制种子链接一键可得) so the user can re-open or share the
   exact same card. State that custom Blender shader graphs are rebuilt with
   GLSL in Three.js; glTF does not carry the node graph, so minor render
   differences are expected. Publish a website or repo only when asked.

随机工坊（shipped UI, do not strip）

The web template ships a randomness panel driven by a seeded PRNG
(assets/web-template/variants.js): 7 palettes × 4 back compositions × 8
sigils × 4 rarities (weighted, with serial ranges) × 6 procedural pets × 5
accessory sets × 5 camera presets × 4 motions × continuous foil/dust/pattern
parameters — all generated on canvas/shaders, zero image credits. Page
controls: 🎲 重掷一张 (new seed, URL-synced), 复制种子链接 (share/reopen the
exact card), camera chips, motion chips, accessory/pet cyclers, dust slider.
Keyboard/drag/flip behaviour and the holographic shader direction stay
unchanged; window.__holo gains variant/seed/reroll/setCamera/playMotion/ cycleAccessory/cyclePet for automation.

声光剧场（shipped UI, do not strip）

assets/web-template/audio.js is a procedural Web Audio engine: zero audio
files, zero binaries (SkillHub rejects them anyway; never add any). One seed
derives a full soundprint (deriveSoundprint: scale × timbre × rhythm × root
× wetness); the card back composition drives a continuous scene bed
(徽记=圣殿 / 星图=星穹 / 纹章=轮转 / 回路=脉冲 / 静默), the signature chord
announces the card, and all interactions speak: flip, rarity reveal, reroll
scan, camera, motions, accessory, six synthesized pet voices (pitch follows
petScale, pan follows petSide, idle chirps while accompanied), a foil slider
that brightens a shimmer band, and a drag "light-blade" friction gain. Default
silent — audio unlocks on the first user gesture (报头「声场」power button);
the preference is remembered but still needs one gesture next visit. Page
controls: spectrum hairline canvas (#spectrum-canvas), soundprint code,
scene chips, pet-voice cycler, SFX chips, volume. window.__holo.audio
(exposes snapshot()/events/stats) and window.__holo.sound serve automation;
references/audio.md is the design law — keep every sound routed through
_log so acceptance can prove it.

Non-obvious invariants (keep intact)

• All image planes are created in XY with an object rotation of X=90°; keep
  that rotation unapplied. Editing the mesh to fake the rotation breaks the
  intended local axes.
• Default core controls: subject scale 1.25 / depth 0.4, background depth
  −0.25, text scale 1 / depth 0. The shared group inputs are 缩放 and
  深度, output 视差效果.
• UV centering plus normal transforms alone cannot make full view-dependent
  parallax: also transform the view direction into the card plane, divide by a
  bounded normal component and offset UV by signed depth (the bundled engine
  does this; preserve it).
• Browser contract uses the material names web_front, web_edge,
  web_back, web_gold. In the viewer, use the canonical card root frame for
  uView, not the exported front mesh's local frame, and restore the exported
  V coordinate exactly once. Test both turn directions to catch smearing or
  inverted parallax.

Resources

• scripts/ — bundled engine (upstream MIT, carries the Blender-listing
  User-Agent fix): ensure_blender.py (official, SHA-256-verified portable
  Blender), build_card.py / export_web.py (editable scene + GLB export),
  generate_typography.py / validate_assets.py (deterministic helpers),
  run_pipeline.py (one-shot pipeline). WorkBuddy glue: prepare_layers.py
  (matting, watermark removal, canvas centering), randomize_card.py (roll a
  seeded variant into card-config.json), verify_card.mjs /
  verify_workshop.mjs / verify_audio.mjs (browser acceptance).
• assets/web-template/ — responsive Three.js viewer + seeded variant engine
  (variants.js) + procedural sonic engine (audio.js; no source artwork, no
  audio files). Visual system: 「光谱标本档案」— paper archive + dark specimen
  vitrine + single indigo accent + spectral hairlines.
• references/design.md — design system: direction, colour/font tokens,
  composition rules, motion, DOM contract and redesign checklist. Read before
  touching the web UI.
• references/audio.md — sonic design system: three-layer structure, seed →
  soundprint derivation, scene/back mapping, synth recipes, autoplay rules and
  mix discipline. Read before touching any sound.
• references/art-direction.md — layered prompts, WorkBuddy image-tool
  semantics (no alpha, watermark, line-art registration) and the
  宠物/配饰/动作/镜头 variety guidance.
• references/randomness.md — seeded variant design: what is rolled, seed
  semantics, reproduction/sharing, window.__holo automation hooks.
• references/config.example.json — card-config schema example (incl.
  optional seed/variant).
• references/verification.md — acceptance checklist + browser tooling.
• references/troubleshooting.md — install/alpha/watermark/ESM pitfalls.
• LICENSE.md — upstream MIT licence (attribution retained).

Never bundle generated cards, downloaded Blender, node_modules, caches, audio
files (all sound is synthesized at runtime), or image-bearing .blend/.glb
files into this skill when distributing it.