SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsxhs-diagnosis

xhs-diagnosis

Marketing & growth

View repository

Updated 10-15 21:41

About this skill

Xiaohongshu Note Diagnosis Engine

A Xiaohongshu note diagnosis and optimization tool based on a 5 AI Agent multi-round debate architecture.

Core Capabilities

Capability Description
Five-Dimensional Diagnosis Content quality · visual presentation · growth strategy · user reaction · overall score
Multi-Agent Debate 4 experts diagnose in parallel → 3 rounds of debate → final review by the judge
Comment Section Simulation Real Xiaohongshu comment style (including arguments/nested replies)
Automatic Rewriting One-click generation of an optimized title + body ready for direct publishing
Baseline Comparison Compare with 874 real hit note data points (5 major categories)
Category Differentiation Food/Fashion/Tech/Travel/Lifestyle each have independent scoring models

How to Use

Trigger Conditions

Load this Skill when the user mentions any of the following needs:

  • "Help me take a look at this Xiaohongshu note"
  • "Diagnose why this note didn't take off"
  • "Help me optimize Xiaohongshu copy/title"
  • "Simulate how the comment section would react"
  • "Give my note a score"

Input Format

The user can provide any of the following forms:

  1. Text description: title + body content + category classification
  2. Screenshot/Image: note cover or full screenshot (requires image_gen or visual capability for analysis)
  3. Link: Xiaohongshu share link (requires parsing to extract content)

Required fields:

  • title: note title
  • content: note body
  • category: category (Food/Fashion/Tech/Travel/Lifestyle/Beauty/Fitness/Home)
  • image_description: cover description (if there is an image, use the visual analysis result; if not, mark "No cover")

Optional fields:

  • tags: currently used tag list
  • publish_time: planned publishing time
  • images_count: number of images

Diagnosis Process (6 Steps)

Step 1: 多模态内容解析
   ├── 标题特征提取（长度、钩子词、emoji）
   ├── 正文特征提取（可读性、信息密度、段落结构）
   └── 图像特征提取（饱和度、文字占比、构图）

Step 2: 品类基线对比
   ├── 与同品类爆款均值对比
   ├── Model A 预评分（确定性公式）
   └── 生成稳定性雷达分

Step 3: 并行专家诊断（4个Agent同时工作）
   ├── ContentAgent  → 内容分析师
   ├── VisualAgent   → 视觉诊断师
   ├── GrowthAgent   → 增长策略师
   └── UserSimAgent  → 用户模拟器

Step 4: 专家辩论（2轮）
   └── 各Agent互相审阅，提出反驳和补充

Step 5: 综合裁判终审
   └── JudgeAgent 汇总全部意见，输出最终报告

Step 6: 组装输出
   └── 五维雷达图 + 问题清单 + 优化方案 + 模拟评论 + 改写版本

Agent Role Definitions

For detailed prompts, see references/prompts/ the files in the directory:

File Agent Responsibility
content_agent.py Content Analyst Title appeal, content structure, information density, readability, keyword strategy
visual_agent.py Visual Diagnostician Color saturation, composition, text overlay, on-camera people, feed competitiveness
growth_agent.py Growth Strategist Content appeal (30%), tag strategy (25%), interaction guidance (20%), publishing time (15%), audience matching (10%)
user_sim_agent.py User Simulator 3 types of user reaction evaluation + 5-8 realistic-style simulated comments (including arguments)
judge_agent.py Comprehensive Judge S/A/B/C/D grading, final ruling, output optimized title + body
debate.py Debate Coordinator Forced rebuttal mechanism to avoid groupthink

Output Format

After the diagnosis is complete, the following complete report structure must be output:

{
  "overall_score": 72,
  "grade": "B",
  "radar_data": {
    "content": 68,
    "visual": 75,
    "growth": 65,
    "user_reaction": 78,
    "overall": 72
  },
  "issues": [
    {"severity": "high", "description": "具体问题+数据支撑", "from_agent": "内容分析师"}
  ],
  "suggestions": [
    {"priority": 1, "description": "具体可执行建议", "expected_impact": "预期效果"}
  ],
  "debate_summary": "共识与分歧总结",
  "optimized_title": "优化后标题（12-20字，含2+钩子）",
  "optimized_content": "优化后正文（小红书排版，300字内）",
  "cover_direction": {
    "layout": "构图建议",
    "color_scheme": "配色建议",
    "text_style": "封面文字建议"
  },
  "simulated_comments": [
    {"username": "昵称", "comment": "评论内容", "sentiment": "positive/negative/neutral", "likes": 数值}
  ]
}

Category Differentiation Parameters

See references/category_baselines.md。

Key differences:

  • Food: Title weight 57.3% (highest), the title determines life or death
  • Fashion: Visual weight 25%, if the images are bad there is basically no hope
  • Travel: Tag weight 31.2%, SEO is more important than content
  • Tech: Interaction potential weight 26.7%, topic discussion level is the key
  • Lifestyle: Tag weight 27.7% + title 40.7%, dual-core driven

Scoring Grade Standards

Grade Score Range Meaning
S 90-100 Excellent in all dimensions, with viral potential
A 75-89 Overall good, with a small room for optimization
B 60-74 Mediocre, with clear directions for improvement
C 40-59 Obvious problems exist, requiring relatively large adjustments
D 0-39 Serious problems, recommended to remake

Execution Guide

When executing as an Agent:

  1. Receive user input: Extract title, content, category from the note content/screenshots provided by the user
  2. Call each Agent prompt in sequence: According to the system prompts in references/prompts/, execute the diagnostic roles one by one
  3. Simulate the debate process: Have each Agent review others' opinions and generate rebuttals and supplements (refer to debate.py)
  4. Judge final review: Use the prompt from judge_agent.py to summarize all opinions and give the final rating and rewrite
  5. Format output: Organize the results into the above JSON format, and also provide a human-friendly Markdown report

Notes

  • All issues must bequantified, vague descriptions are prohibited (❌ "the title is not good enough" ✅ "the title is 14 characters, lower than the category average of 18.3 characters")
  • All suggestions must bedirectly executable(✅ "Title 'XX' → change to 'Tested! 3 Methods for XX' → add numbers + suspense")
  • The comment section must includecontroversy/arguments(about 20%), this is a real ecological characteristic of Xiaohongshu
  • The optimized main text must be within 300 characters, using Xiaohongshu-style formatting (emoji markers + colloquial language + interactive guided ending)
  • When there is no image, mark "No cover image received, the following are inferred suggestions"

Reference Files

  • references/prompts/content_agent.py — Content Analyst Prompt
  • references/prompts/visual_agent.py — Visual Diagnostician Prompt
  • references/prompts/growth_agent.py — Growth Strategist Prompt
  • references/prompts/user_sim_agent.py — User Simulator Prompt
  • references/prompts/judge_agent.py — Comprehensive Judge Prompt
  • references/prompts/debate.py — Debate Stage Prompt
  • references/category_baselines.md — Category baseline data & Model A parameters
  • references/orchestrator_logic.md — Orchestrator process description