/RuiC-card-skill

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.

| layer | file | job | depth | |---|---|---|---| | subject | `assets/subject.png` | the character or hero object, real alpha, composited over the background | `subjectDepth` (0.4) | | effects | `assets/effects.png` (optional) | a pre-cut decorative overlay — petals, sparks, thorn work — that floats **between** the subject and the typography | `effectsDepth` (0.5 forward, or negative to sink it) | | background | `assets/background.png` | the opaque environment behind everything | `backgroundDepth` (−0.25) | | lineart | `assets/lineart.png` | dark contours on white; drives the glowing ink glints that appear inside the subject's silhouette | follows the subject | | text | `assets/text.png` | typography and card furniture. The viewer samples this layer **without parallax**, so a border or frame painted here stays nailed to the card edge | 0 |

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](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](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](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,