write-xiaohongshu

Write Xiaohongshu (小红书: research first, then write, then post)

Quick Start (shortest usage)

When the user gives you a topic, directly run the full process "steps 0→7"; in the end only output:

Title (≤20 characters)
Body (≤1000 characters)
Tags (#xxx #yyy…)
Images (if the user provides them, use the provided images; if not, choose 1~2 yourself, 9:16 preferred)
Publish after user review
Publishing result (if Xiaohongshu MCP is available)

Hard limits (must be followed)

Title:≤ 20 characters
Body:≤ 1000 characters

Counting method: count by "characters" (including spaces, punctuation, line breaks, #topics, etc.). To be safe: keep the title as close to ≤18 as possible, and the body as close to ≤950 as possible, leaving margin for tags/formatting.

Output (must include these blocks)

The output of this skill is divided into three sections:

Analysis summary report(must be given first after completing the analysis in steps 1+2, see the template below for structure)
Final publishable draft(title + body + tags + images)
Publishing result after review

标题：<不超过20字符>
正文：
<不超过1000字符>
标签：#标签1 #标签2 #标签3 ...
配图：<1~2张，9:16优先；给出图片URL或本地路径>

If the publishing interface supports a separate "tags field", then do not pile #tags into the body; put the tags into the publishing parameters; in the end you still need to ensure the body itself is ≤1000.

Overall process (must be done in order)

flowchart TD
  A[检查登录状态] --> B[小红书找Top10图文]
  B --> C[拆标题&正文规律 + 点赞高原因]
  C --> D[抓评论并分析情绪共鸣点]
  D --> E[Firecrawl补背景&事实校验]
  E --> F[写标题/正文/标签 + 字数自检]
  F --> G{用户是否提供图片?}
  G -- 是 --> H[用用户图片]
  G -- 否 --> I[网上找1~2张高清9:16图]
  H --> J[展示内容给用户确认]
  I --> J
  J --> K{用户确认?}
  K -- 确认 --> L[用小红书MCP发布]
  K -- 需修改 --> F

Step 0: First check login status

First call the Xiaohongshu mcp to check the login state; if not logged in, obtain a QR code for the user to log in.

Step 1: Find related "Top 10 by likes" image-text posts on Xiaohongshu, and summarize the patterns

Goal: first find "the most popular way of writing in similar content", then start writing.

Hard threshold (must be executed)

Must successfully obtain at least 10 image-text posts with "complete and usable post details", only then can you enter steps 2/3/4/5/6/7.

Requirements:

Only selectimage-text posts(exclude videos/livestream clips)
Relevance first, then high like count, filter by "most liked"
Filter time "within half a year"
Only use Xiaohongshu mcp's get_specified_post tool to obtain posts filtered by "image-text" and "within half a year", get 20 posts of data (the actual result may be fewer than 20, tolerance 10), do not call other tools of the mcp; the content returned by this tool contains all the information needed.

For each Top post, record these fields (for subsequent analysis):

Title (original text)
Likes/favorites/comments (if available)
Body structure (summarize in 1 line: e.g. "conclusion→list→steps→feedback→interaction")
Key hook (the "attraction point" of the first sentence at the beginning)
Evidence type (experience, comparison, data, screenshots, before-and-after changes…)
Interaction design (questions, polls, comment keywords, giveaways, collection guidance…)

Output a "pattern summary" (directly reusable for writing):

Title patterns (length, sentence structure, keywords, symbols/numbers, emotional words, benefit points)
Body patterns (paragraph rhythm, lists/steps, turns, pitfalls to avoid, closing CTA, tone style)
Reasons for high likes (at least 3, try to ground them in "writing style/information density/points of resonance/actionability/tone style")

Step 2: Based on the comments in the Step 1 data, analyze the reasons for "emotional resonance"

Goal: Find out why viewers "want to comment/want to share/want to save", and summarize 3-6 resonance point conclusions and 1 interaction design suggestion.

Method:

Categorize the "resonance triggers" in the comments, for example:
- Shared pain points (Me too...)
- Identity recognition (office worker/student/mom...)
- Anxiety and solution (Finally there's a way/I'm saved)
- Values (self-discipline, relaxed vibe, cost-effectiveness, minimalism...)
- Contrast/surprise (So this works too)
- Replicability (I can do it too)
- Desire to share (I want to try it too)
- Curiosity (What is this?)
- Resonance point (I'm like this too)

Output:

3-6 "resonance point conclusions" (each conclusion paired with 1-2 original quote fragments/summaries)

Analysis summary report (must be output, then automatically continue with the subsequent process)

After you complete Steps 1+2 (and satisfy "available samples >= 10"),you must first output a summary report according to the structure below(for content style/granularity, refer to the example given by the user), and then in the same round of outputautomatically continue with Steps 3->7。

If the sample threshold is not met: according to the hard threshold requirement in Step 1,stop directly, do not enter Steps 2/3/4/5/6/7, and do not fabricate a report.

Report template:

【分析总结报告】
评论数/互动量排名（从高到低）
1. <标题> - <评论数>（<图文/视频>）
2. ...
...
10. ...
11. ...

标题规律分析
- <规律1>
- <规律2>
- <规律3>
...

内容规律分析
- <规律1>
- <规律2>
...

封面规律分析
- <规律1>
- <规律2>
...

总结
爆款三要素：
1. <要素1>
2. <要素2>
3. <要素3>

互动催化剂：
- <催化剂1>
- <催化剂2>
...

Step 3: Use Firecrawl MCP to fill in background knowledge (and perform fact-checking)

Goal: Avoid empty content, avoid errors/exaggeration, and fill in "credible information".

Method:

Use topic keywords to expand retrieval: concept explanations, common misconceptions, authoritative advice, data (from credible sources as much as possible)
Distill into "5-10 citable background points" (short sentences are fine)
Downgrade claims that are easily questioned (efficacy, benefits, comparative conclusions):
- Change them to "my personal experience/suitable for some people/may be effective"
- Or add conditions (applicable population, prerequisites)

Output a "background knowledge card":

Key conclusions (5-10 items)
Risks/precautions (2-4 items)

Step 4: Based on the above analysis, write the title + body + tags (and force a word count self-check)

Writing requirements (leaning toward Xiaohongshu, but don't be pretentious):

Colloquial, short sentences, line breaks every 1-3 sentences
High information density: less empty talk, more "steps/checklists you can follow"
Can appropriately use emoji (to enhance rhythm/emotion), but do not pile them on
The ending should be natural, like something a real user wrote, not a "marketing account tone"

De-AI-flavor prompt words (must be executed)

Write more "I/At the time I/I discovered/the pitfalls I stepped in", and use less "everyone/users/audience/suggestions are as follows"
Give 2–3 specific details: time point, scene, before-and-after comparison, a specific line of dialogue/a shot/a small action
Imperfection is allowed: use "I feel/I guess/maybe/not sure but…" (don't pretend to be an authority)
Fewer template phrases: avoid "in summary/in conclusion/therefore/firstly secondly finally/must not miss/hurry"
Like chatting with a friend: short sentences, line breaks, occasional colloquial words (but don't pile them up)

Word count gate (must be done last)：

Title character count > 20: delete redundant words, remove subtitles, shorten sentences, remove unnecessary symbols
Body character count > 1000: first delete repetition and catchphrases, then merge sentences, then compress steps/reduce cases
Finally output only the compliant version

Step 5: If the user hasn't provided images, find 1–2 high-definition images online yourself (9:16 preferred)

Requirements:

1–2 are enough, high-definition, clean, strongly related to the topic
Aspect ratio 9:16 preferred (vertical image)
Only choosecommercially usable/no copyright riskpreferred sources (e.g., Pexels / Unsplash, etc.)
When outputting, provide: image URL (and optionally: local download path)

Step 6: Show the content to the user for confirmation (Review)

Goal: let the user confirm the content before publishing to ensure it meets expectations.

Must execute：

After completing steps 4+5,you must first show the complete content to the user for confirmation, in the following format:

【待发布内容预览】

标题：<标题内容>（<字符数>/20）
正文：
<正文内容>
（<字符数>/1000）

标签：#标签1 #标签2 #标签3 ...

配图：
- 图片1：<URL或路径>
- 图片2：<URL或路径>（如有）

---
请确认：
- 输入"确认"或"发布"：将进入发布流程
- 输入"修改"或具体修改意见：将根据您的意见调整内容
- 输入"取消"：终止发布流程

Wait for user confirmation：

If the user confirms: proceed to step 7 to publish
If the user requests modifications: modify the content according to the user's feedback and show it again for confirmation (return to step 6)
If the user cancels: terminate the process, do not publish

Note: before the user explicitly confirms,you must not automatically proceed to the publishing step。

Step 7: Publish using the Xiaohongshu MCP

When publishing (adapt to MCP capabilities, do as much as possible):

Upload 1–2 images (vertical images preferred)
Set the title, body, and tags
If supported: first create a draft → echo the draft information → then publish

After successful publishing, output:

Published link/ID (if available)
Publishing parameter summary (title/tags/number of images)

Fallback strategy (when tools are unavailable)

If the current environment does not have Firecrawl MCP or Xiaohongshu MCP:

Still complete "analysis → writing → image links" according to the process, but do not auto-publish
List the information the user needs to supplement as the shortest checklist (e.g., provide 3–5 benchmark post links, or authorization/API configuration)

Examples

Example 1: The user only gives a topic

User input:

Write a Xiaohongshu image-and-text post: the topic is "quick dinners for office workers," suitable for one person, with a budget of under 20 yuan.

What you need to do:

Use the topic keywords to find "image-and-text posts" on Xiaohongshu and summarize the title/body patterns
Capture the comments on these posts and extract resonance points and interactive questions
Use Firecrawl to supplement background related to "health/meal prep/safety"
Output the title + body + tags, and perform a word count self-check
If there are no images, find 1–2 high-definition 9:16 vertical images
Show the content to the user for confirmation
After the user confirms, publish using the Xiaohongshu MCP (if available)

Example 2: The user provides images