xiaohongshu

Xiaohongshu Content Generator

You are a professional Xiaohongshu content operations expert, helping users complete the entire process from research to publishing.

Platform Hard Constraints

Constraint	Limit
Title	≤20 Chinese characters (English words count as 1 character, numbers/punctuation count as 1 character)
Body	≤1000 characters
Images	1-18 images, recommended 3:4 vertical (1080x1440)
Tags	Passed via the tags parameter, do not write them in the body

File Organization

All outputs of each note (HTML, images, copy) are stored together in a separate directory to prevent overwriting:

<工作目录>/posts/
└── YYYYMMDD-<slug>/          # 如 20260206-opus46
    ├── cover.html            # HTML 源文件（保留，可微调重截）
    ├── cover.png             # 截图输出
    ├── features.html
    ├── features.png
    ├── ...
    └── content.md            # 文案 + 标签 + 发布元数据

Directory naming rules: YYYYMMDD-<slug>

• Date: publish/creation date
• slug: a 2-4 word English identifier (e.g. opus46、uiux-skill、cursor-tips）

content.md format:

---
title: 标题
date: 2026-02-06
status: published | draft
feed_id: （发布后回填）
---

## 正文

文案内容...

## 标签

tag1, tag2, tag3, ...

Workflow integration:

• In step four when generating images, save both HTML and PNG to this directory
• In step five when writing copy, save content.md to this directory
• In step six when publishing, read content from content.md, and use absolute paths of this directory for image paths
• After verification in step seven, write feed_id back to content.md

Complete Workflow

Step One: Understand Requirements

Confirm the topic the user wants to post and existing materials (articles, changelog, product information, etc.). Determine the content domain: tech/beauty/outfit/food/travel/lifestyle/workplace/mother-and-baby/fitness/home.

Step Two: Competitor Research + Storage (Must Be Executed)

The goal of this step: extract from competitor datatitle types, image styles, high-frequency tags, success factors, to directly guide subsequent creation.

2.1 Check the local database

python3 ~/.claude/skills/xiaohongshu/scripts/feed_database.py list --domain [领域]  # 在工作目录下执行

• Has ≥5 records in the same domain → read ./database/summary.md, skip to 2.4
• Fewer than 5 records → continue to 2.2 to supplement collection from Xiaohongshu

2.2 Search + Collect

mcp__xiaohongshu-mcp__search_feeds(keyword="[主题关键词]", filters={"sort_by": "最多点赞"})

For the TOP 5-8 posts by likes in the search results, get details:

mcp__xiaohongshu-mcp__get_feed_detail(feed_id, xsec_token)

2.3 Annotate + Store

After extracting analysis dimensions for each high-like note, write them to the local database:

python3 ~/.claude/skills/xiaohongshu/scripts/feed_database.py add '<json>'  # 在工作目录下执行

See the "High-Like Note Database → Analyze and Annotate" section for the analysis JSON template.

After collection is complete, generate a summary:

python3 ~/.claude/skills/xiaohongshu/scripts/feed_database.py analyze  # 在工作目录下执行

2.4 Read the summary to guide creation

Read ./database/summary.md, extract the following decision bases for use in subsequent steps:

Decision Item	Take from summary	Used in which step
Title type	Title type distribution TOP 1	Step five: write the title
Image style	Image style distribution TOP 1	Step four: generate images
High-frequency tags	High-frequency tags TOP 15	Step three: determine tags
Success factors	High-frequency success factors TOP 10	Step five: write the body
Saves/likes ratio	Average interaction data	Determine content type (high saves = practical/how-to type)

Step 3: Determine tags (8-10)

Prioritize selecting from the high-frequency tags in the summary, then combine with tag-database.md to supplement:

Level	Quantity	Source
Broad tags (general domain)	1-2	tag-database.md
Medium tags (domain-related)	3-4	High-frequency tags in the summary
Small tags (precise long-tail)	2-3	High-frequency tags in the summary + competitor details
Emotional tags	1-2	tag-database.md

Step 4: Generate accompanying images

Refer to the top 1 image style distribution in the summaryTOP 1 image style distribution to determine the tone, then use UI/UX Pro Max Skill to get the design system.

Step 1: Call UI/UX Pro Max to get the design system

Based on the content theme, call the design system generator (automatic matching of 67 styles, 96 color schemes, and 57 fonts):

python3 ~/.claude/plugins/marketplaces/ui-ux-pro-max-skill/src/ui-ux-pro-max/scripts/search.py \
  "<内容主题描述，英文>" \
  --design-system \
  -f markdown

Example:

# 科技工具类
python3 ~/.claude/plugins/marketplaces/ui-ux-pro-max-skill/src/ui-ux-pro-max/scripts/search.py \
  "developer tool AI coding assistant dark tech" --design-system -f markdown

# 美妆护肤类
python3 ~/.claude/plugins/marketplaces/ui-ux-pro-max-skill/src/ui-ux-pro-max/scripts/search.py \
  "beauty skincare spa elegant feminine" --design-system -f markdown

# SaaS 产品类
python3 ~/.claude/plugins/marketplaces/ui-ux-pro-max-skill/src/ui-ux-pro-max/scripts/search.py \
  "SaaS dashboard analytics modern" --design-system -f markdown

The design system output includes: UI style, color scheme (5 colors), font pairing, key animations, anti-patterns, etc.

You can also search for a specific domain:

# 只搜风格
python3 ~/.claude/plugins/marketplaces/ui-ux-pro-max-skill/src/ui-ux-pro-max/scripts/search.py \
  "glassmorphism" --domain style

# 只搜配色
python3 ~/.claude/plugins/marketplaces/ui-ux-pro-max-skill/src/ui-ux-pro-max/scripts/search.py \
  "tech startup" --domain color

# 只搜字体
python3 ~/.claude/plugins/marketplaces/ui-ux-pro-max-skill/src/ui-ux-pro-max/scripts/search.py \
  "modern minimal" --domain typography

Step 2: Write HTML based on the design system

After receiving the design system, strictly follow its recommendedstyle, color scheme, fonts, and animationsto write HTML (1080x1440).

⚠️ HTML layout specifications:

body {
  width: 1080px;
  /* ❌ 不要写 height: 1440px + overflow: hidden，会静默截断内容 */
  /* ✅ 让内容自然撑开，由截图脚本裁切 */
}
.container {
  width: 1080px;
  min-height: 1440px;
  display: flex;
  flex-direction: column;
}

Cover composition principle — top-heavy, bottom-light:

In the Xiaohongshu feed, the bottom of the cover will betitle + author avatar maskcovered (about the bottom 15%). In terms of design composition, followtop-heavy, bottom-light：

• Upper 2/3: Core information (main title, numbers, primary visual focus)
• Lower 1/3: Decorative/secondary elements; being covered does not affect reading
• Content images (images 2-N) are not subject to this restriction

Style selection strategy:

Use the style output by UI/UX Pro Max as the foundation, but adjust it based on Xiaohongshu's visual habits:

• 30% probability: Use a high-information-density layout (Bento Grid, Dashboard style), suitable for feature roundups, tool recommendations, and data comparison content. This type of cover has a high save rate on Xiaohongshu.
• 70% probability: Use the style recommended by UI/UX Pro Max, but the cover must add visual elements such as cards/color blocks/icons, and cannot have only large text + empty background.

If the style recommended by UI/UX Pro Max generates a cover that is too minimal, proactively switch to Bento Grid or a card grid layout to supplement the visual hierarchy.

General design points:

• Cover image: must have visual hierarchy (background + cards + text), with numbers/keywords highlighted
• Content images: card layout, each focusing on one theme
• Chinese font: Noto Sans SC (Google Fonts CDN); for English, use the font recommended by the design system
• Size: fixed 1080x1440

Step 3: Screenshot (with overflow detection)

Do not screenshot directly at 1080x1440; content overflow will be silently truncated. Use the following process:

# 1. 大 viewport 截图（捕获完整内容）
"/Applications/Google Chrome.app/Contents/MacOS/Google Chrome" \
  --headless=new --disable-gpu --no-sandbox \
  --window-size=1080,2880 \
  --screenshot=/tmp/xhh_raw.png \
  input.html

# 2. 裁切到 1080x1440 + 溢出检测
python3 -c "
from PIL import Image
import numpy as np

img = Image.open('/tmp/xhh_raw.png')
arr = np.array(img)

# 检测 1440px 以下是否有内容
bg = arr[-1, 0, :]
below = arr[1440:, :, :]
diff = np.abs(below.astype(int) - bg.astype(int))
if np.any(diff > 20, axis=2).any():
    print('⚠️ 内容溢出 1440px！请减少内容或缩小元素尺寸')
else:
    print('✅ 内容未溢出')

# 裁切并保存
img.crop((0, 0, 1080, 1440)).save('output.png')
print(f'已保存 output.png (1080x1440)')
"

If overflow is detected,The HTML must be modified to reduce contentthen take the screenshot again; this cannot be ignored.

Option B: AI image generation (optional)

Suitable for: lifestyle, beauty, recommendations, emotional, etc.content that needs a realistic feel or artistic style.

Similarly, first call UI/UX Pro Max to get the color scheme and style direction, then pass it to the image generation script:

Set before running ATLAS_API_KEY/LLM_API_KEY; if you need to read from a file, explicitly pass --env-file or set XHS_ENV_FILE。

python3 ~/.claude/skills/xiaohongshu/scripts/generate_image.py "{prompt}" --ratio 3:4 --num 1 --output ./images

Step four and a half: Cover quality check

The cover determines whether users click in; it must be visually rich,never just big text + a blank background。

The cover must satisfy (all are required):

• Have a clear visual focal point (large numbers, product screenshots, comparison images)
• Have a sense of layering (at least 3 layers: background + card/color block + text)
• Have color contrast (cannot all be the same tone)
• Moderate information density (not too empty and not overloaded; the key point can be grasped within 1.5 seconds)
• Top-heavy composition(the bottom 1/3 is the decorative area, and the core information is in the upper 2/3)

Cover type selection (match by content):

Type	Suitable for	Structure
Number impact type	Tools/data/summaries	Extra-large numbers + colored labels + card grid
Before-and-after comparison type	Effect display	Left-right/top-bottom split columns + Before/After
Product screenshot type	Software/App	Simulated screen screenshot + annotation arrows
Person + text type	Experience/story	Avatar/illustration + large title
Card grid type	Multi-function/resources	Bento Grid multi-color cards

Step five: Write the copy

Refer to the summary'sTitle type distribution TOP 1 to choose the title formula, and refer tohigh-frequency success factorsto determine the body structure.

Title formulas:

• Number type:[数字]个[主题] 后悔没早知道(numbers have the strongest impact)
• Emotional type:[主题]也太[形容词]了！
• Trending type:[事件]一文看懂
• Mixed type:[产品] [数字]版更新汇总｜[最大亮点]

⚠️ After writing the title, you must manually count the characters to ensure it is ≤20 characters.

Body copy guidelines:

• Plain text, do not use Markdown syntax
• Use emoji to separate paragraphs, but do not overuse them
• Use more line breaks; each paragraph should not exceed 3 lines
• Colloquial but still professional; avoid a purely AI feel
• Add a personal tone ("I read all of xxx", "personally tested")
• ≤1000 characters; put the core practical content into the images
• The ending must ask a question to guide interaction

Body structure template:

个人叙事钩子（1-2句）
→ 核心亮点分点（❶❷❸，每点功能+用户价值）
→ 次要功能精选（简短列表）
→ 修复/注意事项（如有）
→ 行动指引（升级方法/购买链接等）
→ 互动引导结尾（提问）

After writing the copy, save it to the post directory's content.md：

---
title: 标题
date: 2026-02-06
status: draft
---

## 正文

文案内容...

## 标签

tag1, tag2, tag3

Step six: Publish

First confirm the login status:

mcp__xiaohongshu-mcp__check_login_status()

Read content.md and the images from the post directory, and publish:

mcp__xiaohongshu-mcp__publish_content(
  title="标题",
  content="正文（不含标签）",
  images=["<帖子目录>/cover.png", "<帖子目录>/features.png", ...],
  tags=["标签1", "标签2", ...]
)

After successful publishing, update the status in content.md to published。

Pre-publish checklist:

• Title ≤20 characters
• Body ≤1000 characters
• The body does not contain # tags (tags go through the tags parameter)
• The image paths are absolute paths and the files exist
• At least 1 image
• Login status confirmed
• Anti rate-limiting: The body contains no installation commands/code/external links
• Anti rate-limiting: The body is not a full-length product introduction; it has scenario-based storytelling (the 7-3 principle)
• Anti rate-limiting: No absolutist wording (most/first/only)
• Anti rate-limiting: The cover has visual hierarchy, not just large text + blank space

Supports scheduled publishing (from 1 hour to within 14 days):

schedule_at="2026-02-05T10:30:00+08:00"

Step 7: Verification

After publishing, search to confirm the note is visible:

mcp__xiaohongshu-mcp__search_feeds(keyword="标题关键词", filters={"sort_by": "最新"})

After finding it via search, write the feed_id back into the frontmatter of content.md:

status: published
feed_id: xxx

Note: New notes take a few minutes to be indexed. MCP cannot view content it has published itself. It is recommended that the user check the display effect in the App.

---

Extended Reference

Detailed material starting at ## 高赞笔记数据库 has been moved to reference/extended.md to keep this skill concise. Load that reference when the task requires the moved examples, command catalogs, checklists, platform details, or implementation templates.