| name | rednote-card-generator |
| description | Use when you need to generate Xiaohongshu (小红书 / RedNote) carousel card HTML layouts. Given a piece of text content, this skill produces a complete, self-contained HTML document containing multiple vertical 450×600px cards styled for the RedNote / Xiaohongshu platform. |
RedNote Card Generator
Overview
Use rednote-card-generator to turn any text content into a beautifully styled Xiaohongshu (小红书) carousel post. The output is a single, self-contained HTML string that renders multiple vertical cards (each 450×600 px) stacked vertically with a 20 px gap.
The skill calls an OpenAI-compatible API and streams the HTML response back. Provider/model are controlled by environment variables (LLM_API_KEY, LLM_BASE_URL, LLM_MODEL).
Inputs to collect
| Field | Required | Description |
|---|
content | Yes | The text content to turn into RedNote cards. Can be plain text or light HTML. |
colorScheme | No | One of: auto (default), warm-brown, sunset, ocean, forest, lavender, monochrome. Controls the dominant palette. |
llmApiKey | Yes | API key for your OpenAI-compatible provider. |
Output
Returns a JSON object:
{
"html": "<!DOCTYPE html>...",
"cardCount": 5
}
html: Complete self-contained HTML document. Render this inside an iframe or a sandboxed div.cardCount: Estimated number of cards generated.
Actions
Generate cards
{
"action": "generate",
"content": "今天分享三个提升专注力的方法...",
"colorScheme": "warm-brown",
"llmApiKey": "your-key"
}
Color scheme options
| id | Description |
|---|
auto | AI picks colors based on content mood |
warm-brown | Primary #3E2723, background #FFFCF8 — cozy editorial |
sunset | Primary #FF6B6B, secondary #FFD93D — vibrant warm |
ocean | Primary #4A90E2, secondary #50C9CE — fresh cool |
forest | Primary #2D6A4F, secondary #52B788 — natural green |
lavender | Primary #9D84B7, secondary #C8B6E2 — soft purple |
monochrome | Primary #2C3E50, accent #E74C3C — clean minimal |
Card layout specification
The generated HTML follows these rules so downstream renderers can reliably parse it:
- Outer container:
div with display:flex; flex-direction:column; gap:20px
- Each card:
div with width:450px; height:600px; position:relative; overflow:hidden
- Card backgrounds: warm paper tones —
#FDF8F3, #F5EFE6, #FFFBF5, #F0EBE3 — varied per card
- Top accent strip:
div with height:8px; width:100% in a warm accent color
- Typography:
- Card title / section number: 30–38 px,
font-weight:700, Playfair Display or Space Grotesk
- Body text: 15–17 px,
line-height:1.75, Noto Sans SC or Inter
- Labels / captions: 12–13 px, muted color
- Text colors: coffee-brown tones
#3E2723, #664A42, #5D4037
- Accent shapes: left-border bars (6 px wide), bullet circles (10×10 px), number badges (32×32 px, border-radius 16 px)
- NO emojis in generated content
- Fonts imported via
@import inside <style> — never via <link> tags
Card types
The AI selects the best card type for each section of the content:
| Card type | When to use |
|---|
| Title card (first) | Large decorative title + subtitle + author/tag label |
| Section header | Bold large number (01, 02…) + section heading + 1-sentence teaser |
| Text block | Body paragraphs with colored left-accent bar |
| Bullet list | Key points as rows with filled-circle bullet shapes |
| Numbered steps | Step-by-step instructions with rounded number badges |
| Quote / highlight | One key sentence enlarged (28–34 px), centered, with decorative quote marks |
| Summary (last) | Recap of core message + thin decorative bottom strip |
System prompt used
When calling the LLM, the following system instruction is used:
You are an expert graphic designer who generates production-ready HTML layouts with embedded CSS. Your output is a complete, self-contained HTML document that can be rendered directly in a browser. Output ONLY valid HTML. No markdown, no code fences, no explanation. Include a <style> tag with all CSS. Do NOT use Noto Sans SC. Do NOT set margin or padding on html or body tags. Do NOT use global CSS resets.
User prompt template
The user message sent to the LLM is constructed as follows:
This is a Xiaohongshu (小红书) carousel post. Generate MULTIPLE vertical cards (each 450×600px) stacked vertically with 20px gap between them.
CONTENT FIDELITY — INCLUDE EVERYTHING:
- Cover ALL of the user's text. Do NOT omit, summarize, or skip any sentence.
- Distribute content logically across cards: one topic or section per card.
- Generate as many cards as needed to fit all content (typically 3–10 cards).
Content to include:
"<USER_CONTENT>"
<COLOR_INSTRUCTION>
Return ONLY valid HTML with embedded CSS. No markdown, no code fences, no explanation.
Where <COLOR_INSTRUCTION> is:
- If
colorScheme is auto: Analyze the content and choose bold, contextually appropriate colors and modern typography.
- Otherwise:
Color Palette — foundation colors are: "<primary>" (primary) and "<secondary>" (secondary). Use them as the dominant palette with tints and shades for depth. Do NOT apply them flatly.
API call details
POST <LLM_BASE_URL>/chat/completions
Authorization: Bearer <LLM_API_KEY>
Content-Type: application/json
{
"model": "<LLM_MODEL>",
"messages": [
{ "role": "system", "content": "<SYSTEM_PROMPT>" },
{ "role": "user", "content": "<USER_PROMPT>" }
],
"stream": true,
"temperature": 0.4
}
Default model example: anthropic/claude-3.5-sonnet (customizable via LLM_MODEL)
Post-processing steps
After receiving the full streamed HTML, apply these cleanup steps in order:
- Strip leading/trailing markdown fences (
```html … ```) if present.
- Fix broken
<img> tags — collapse internal whitespace/newlines to a single space.
- If the output does not contain
<html or <!DOCTYPE, wrap it:
<!DOCTYPE html><html lang="zh-CN"><head><meta charset="UTF-8"><title>RedNote</title></head><body>…</body></html>
- Count the number of card
div elements to populate cardCount in the response.
Ideas to try
- Paste a WeChat article and get a ready-to-publish 小红书 carousel.
- Summarize a product review into a 5-card visual post.
- Turn meeting notes into a step-by-step visual guide.
- Convert a recipe into numbered-step cards with a title cover.
