SuperAPI AI Directory

xhs-writer-factory

Marketing & growth

View repository

Updated 10-15 20:30

About this skill

Xiaohongshu Creation Skill Factory

A meta-skill that extracts style features from 1-5 satisfactory notes and trains an exclusive Xiaohongshu creation Skill. Core principle: all analysis work is LLM-driven and does not rely on code execution


Core Positioning

This Skill is a "factory" — it receives 1-5 satisfactory Xiaohongshu notes from the user (images + copy) and outputs an exclusive creation Skill through a 4-step process.

Finished Product Form:

用户输入 → 选题模块 → 分析模块 → 撰写模块 → 绘画提示词模块

Output Directory:./output/{品牌名}-xhs/


4-Step Process Overview

Step 1: 内容提取(./extracted_notes/)
Step 2: 分析归纳(./analysis/)
Step 3: Skill组装(./output/{品牌名}-xhs/)
Step 4: 双维审核 + 输出确认

Step 1: Content Extraction

Task: Extract all content elements from each note one by one, output to the ./extracted_notes/ directory

Detailed Execution Manual: See references/phase-guides.md#step-1

Directory Structure:

./extracted_notes/
  note_1/
    - title.md
    - image_1_analysis.md
    - image_2_analysis.md
    - ...
    - copy_structure.md
    - image_prompts.md
  note_2/
  ...

Execution Content:

  1. Title Extraction: structure type, word count, suspense technique, core elements
  2. Image 7-Module Analysis: purpose judgment + copy reverse-engineering for each image
  3. Body Structure: paragraph function, opening hook, ending CTA
  4. Drawing Prompt Reverse-Engineering: complete 7-module prompt for each image

✋ Confirmation Point: After all of Step 1 is complete, display a summary of the extraction results for each note and ask the user to confirm whether they are accurate


Step 2: Analysis and Synthesis

Input:./extracted_notes/ all content

Output to the ./analysis/ directory:

./analysis/
  - style_guide.md        # 写作风格指南
  - topic_analysis.md     # 选题分析指南
  - image_style.md        # 绘画风格指南

Detailed Execution Manual: See references/phase-guides.md#step-2

Execution Content:

  1. Writing Style Synthesis(11 dimensions): brand information/title style/body structure/mandatory paragraph opening phrases/metaphor system/register features/fixed ending/topic tags/prohibition list/image preferences/cover and end-page structure
  2. Topic Analysis: hot topic types/entry angles/keyword library/formula templates
  3. Drawing Style: visual type/color scheme system/composition rules/element symbol library/image purpose types/quantity combination rules

⚠️ Brand name, brand slogan, and fixed ending templateare thehighest-priority extraction items

✋ Confirmation Point: After all of Step 2 is complete, display the three synthesis documents and focus on confirming whether the "brand information" is accurate


Step 3: Skill Assembly

Input:./analysis/ three synthesis documents + the "brand information" extracted in Step 2

Output Directory:./output/{品牌名}-xhs/

Detailed Execution Manual: See references/phase-guides.md#step-3

Fixed Output Structure:

./output/{品牌名}-xhs/
  ├── SKILL.md
  ├── references/
  │   ├── style_guide.md          # 写作风格指南(个性化提取)
  │   ├── topic_analysis.md        # 选题分析指南(个性化提取)
  │   ├── image_style.md           # 绘画风格指南(个性化提取)
  │   ├── 固化模块.md              # 合规红线(固化资源)
  │   ├── creativity_rules.md      # 爆款规律库(固化资源)
  │   └── image_styles.md          # 图片风格预设(固化资源)
  └── assets/
      └── prompt_templates.md      # 提示词模板

Execution Content:

  1. Generate the SKILL.md main file (brand name has highest priority)
  2. Generate the 3 synthesis files under references/
  3. Generate assets/prompt_templates.md
  4. Solidified resource copying(must be executed):
    • references/compliance.md → compliance red lines
    • references/creativity_rules.md → viral formula library
    • references/image_styles.md → image style presets
  5. Image style preset selection (user provides images → style extraction anchoring; no images → choose 1 of 6)

✋ Confirmation Point: After Step 3 is complete, list the generated file structure and ask the user to confirm whether they are satisfied


Step 4: Dual-dimension review + output confirmation

Task: Review the quality of the finished Skill

Detailed execution manual: See references/phase-guides.md#step-4

4.1 File structure completeness check

File Minimum lines Must exist
SKILL.md 200 lines Required
references/style_guide.md 100 lines Required
references/topic_analysis.md 50 lines Required
references/image_style.md 50 lines Required
assets/prompt_templates.md 30 lines Required

Brand name consistency check: Whether the brand name is consistent in the name field of SKILL.md + all files

4.2 Format review

  • SKILL.md has frontmatter (name + description)
  • The process is clear (user input → each Phase → output)
  • No unreplaced placeholders (all {} variables have actual content)
  • The references directory is referenced (in the main text)
  • The mandatory paragraph opening phrases are written into the writing module
  • The metaphor system is written into the writing module
  • The prohibition list is written into the writing module

4.3 Content testing

  1. Randomly select 1 article from the original input cases
  2. Infer the topic theme of that case
  3. Use that theme to invoke the finished Skill for generation
  4. Compare the output with the original case (brand name/completeness/title style/main text structure/paragraph opening phrases/metaphor system/art style consistency)

Pass criteria: ≥80%, at most 3 rounds of looping

4.4 Output confirmation

  1. Ask the user for the finished Skill output directory (default:./output/{品牌名}-xhs/)
  2. Check whether a Skill with the same name already exists in the target directory → ask whether to overwrite or rename
  3. Complete the final delivery

✋ Final confirmation point: After passing the dual-dimension review, confirm the output directory and installation method, and complete the delivery.


Version history

Version Date Optimization content
v1.0 - Initial version
v9.0 2026-03-26 Established the 4-step process
v10.0 2026-03-26 Added 8-dimension extraction, fixed resource copying, file structure check, and output directory confirmation
v11.0 2026-03-27 Refactored the three-layer loading architecture, compressed SKILL.md to ~300 lines, and deleted redundant files