Skip to content
Back to search

find-influencer-skill

跨平台达人发现与内容评估工作流 Skill

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/guzhezao--find-influencer-skill/badge.svg)](/listings/guzhezao--find-influencer-skill)

HTML

<a href="/listings/guzhezao--find-influencer-skill"><img src="/api/listings/guzhezao--find-influencer-skill/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

# 1. Clone Skill
git clone https://github.com/cool111111/find-influencer-skill.git \
  ~/.your-agent/skills/find-influencer

# 2. Check dependencies
bash ~/.your-agent/skills/find-influencer/scripts/setup_check.sh

# 3. 在 Agent 里说:
#    "帮我找 10 个小红书上做 XXX 的博主"

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

SKILL.md · Manifest · 492 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. ---
  2. name: find-influencer
  3. description: |
  4. Search and evaluate high-potential content creators across platforms (Xiaohongshu, Douyin, Bilibili, etc.).
  5. Interactive workflow: collect requirements -> search -> filter by data -> deep content analysis -> report.
  6. Use when: user wants to find creators, bloggers, KOLs, influencers, or evaluate content quality.
  7. Triggers: find influencer, find creators, find bloggers, search KOL, find influencers, talent scout, creator analysis, find-influencer, 找达人, 搜索博主, 达人检索, 博主分析.
  8. ---
  9. # Creator Discovery & Analysis
  10. Multi-platform creator discovery tool for brand collaboration. Finds high-potential creators through multi-layer keyword search (exact product + competitor + broader domain), filters by recent engagement data (not just follower count), and evaluates subjective brand fit (content style, AI/tech experience, competitor history, visual quality, audience match).
  11. ARGUMENTS: User's initial request (platform, content direction, etc.)
  12. ## Phase 1: Requirements Gathering
  13. Use AskUserQuestion to collect:
  14. 1. **Target Platform(s)**: Xiaohongshu / Douyin / Bilibili / YouTube / TikTok / X(Twitter) / Instagram / multiple
  15. 2. **Product/Brand**: What product or brand is this collaboration for? e.g., ChatGPT, Claude, Gemini
  16. 3. **Content Direction**: The specific topic AND broader category. e.g., "ProductX (broader: AI agent, AI工具)"
  17. 4. **Collaboration Goal**: What kind of content do you want the creator to produce? e.g., 产品测评, 教程, 创意展示
  18. 5. **Tone Preference**: What style fits the brand? e.g., 专业但不枯燥, 轻松科普, 极客硬核, 创意炫酷
  19. 6. **Follower Range**: default 5,000 - 500,000 (soft reference, not hard cutoff -- great recent data can override low followers)
  20. 7. **Recent Data Priority**: default 1 month. Minimum acceptable recent post engagement (likes/saves/views)
  21. 8. **Known Competitors**: List competitor products so we can check if creators have collaborated with them. e.g., Coze, Dify, FastGPT
  22. 9. **Number of Results**: default 5-10 creators
  23. If ARGUMENTS already contain these details, skip redundant questions.
  24. ## Phase 2: Content Search
  25. ### Tool Selection
  26. | Platform | Primary Tool | Fallback |
  27. |----------|-------------|----------|
  28. | Xiaohongshu | `opencli xiaohongshu search` | Playwright browser |
  29. | Bilibili | `opencli bilibili search` | Playwright browser |
  30. | Douyin | Playwright browser | - |
  31. | YouTube | `opencli youtube search` | Playwright browser |
  32. | TikTok | `opencli tiktok search` | Playwright browser |
  33. | X (Twitter) | `opencli twitter search` | Playwright browser |
  34. | Instagram | `opencli instagram search` | Playwright browser |
  35. ### Search Strategy: Multi-Layer Keyword Expansion
  36. Don't just search the exact product name -- use a 3-layer keyword strategy to find both vertical and adjacent creators:
  37. **Layer 1: Exact product/brand keywords** (find creators already covering your product)
  38. - Direct product name and variations
  39. - Example for ProductX: "ProductX", "productx教程", "productx测评"
  40. **Layer 2: Competitor & category keywords** (find creators in the same space)
  41. - Competitor product names from Phase 1
  42. - Category terms that your product belongs to
  43. - Example: "Coze教程", "AI agent平台", "AI工作流搭建", "Dify教程"
  44. **Layer 3: Broader domain keywords** (find quality AI creators who could pivot to your product)
  45. - Broader topic keywords in the same domain
  46. - Adjacent content areas where your audience overlaps
  47. - Example: "AI工具推荐", "AI效率提升", "AI科技测评", "AI产品体验"
  48. **Execution:**
  49. 1. Generate 2-3 keywords per layer (6-9 total), search each with `sort=popularity_descending` when available
  50. 2. Extract: title, author name, author ID/URL, likes count, saves count (收藏), post date
  51. 3. **Tag each result with its source layer** -- Layer 1 hits are highest relevance, Layer 3 are expansion candidates
  52. 4. **IMPORTANT**: The author field from search results often contains a date suffix (e.g., "博主名02-05", "博主名2天前"). Parse this date and use it to pre-filter -- only keep posts from within the recency window (default: 1 month). A high-likes post from 6 months ago is NOT evidence of current quality.
  53. 5. Filter to recent posts (within 1 month by default) with 赞藏数 (likes + saves) meeting user's threshold
  54. 6. Deduplicate by author -- prefer candidates with MULTIPLE recent high-engagement posts over one-hit wonders
  55. 7. When deduplicating, preserve the **highest-relevance layer tag** (if a creator appears in both Layer 1 and Layer 3, tag as Layer 1)
  56. ### Platform-Specific Search
  57. **Xiaohongshu via opencli:**
  58. ```bash
  59. opencli xiaohongshu search "<keyword>" --limit 20 -f json
  60. ```
  61. If blocked by login wall, use Playwright:
  62. ```
  63. Navigate: https://www.xiaohongshu.com/search_result?keyword=<encoded>&type=1&sort=popularity_descending
  64. Close login modal if present, then extract via browser_evaluate
  65. ```
  66. **Bilibili via opencli:**
  67. ```bash
  68. opencli bilibili search --keyword "<keyword>" --limit 20 -f json
  69. ```
  70. **TikTok via opencli:**
  71. ```bash
  72. opencli tiktok search "<keyword>" --limit 20 -f json
  73. # Get profile stats
  74. opencli tiktok profile <username> -f json
  75. # Get recent videos
  76. opencli tiktok user <username> --limit 20 -f json
  77. ```
  78. **X (Twitter) via opencli:**
  79. ```bash
  80. opencli twitter search "<keyword>" --limit 20 -f json
  81. # Get profile stats (followers, bio)
  82. opencli twitter profile <username> -f json
  83. ```
  84. **Instagram via opencli:**
  85. ```bash
  86. opencli instagram search "<keyword>" --limit 20 -f json
  87. # Get profile stats
  88. opencli instagram profile <username> -f json
  89. # Get recent posts
  90. opencli instagram user <username> --limit 20 -f json
  91. ```
  92. **YouTube via opencli:**
  93. ```bash
  94. opencli youtube search --query "<keyword>" --limit 20 -f json
  95. # Get video metadata (views, likes)
  96. opencli youtube video "<url>" -f json
  97. ```
  98. For YouTube channel subscriber count, use Playwright to visit `https://www.youtube.com/@<handle>`.
  99. Refer to `guides/platform-selectors.md` for DOM selectors and JS extraction code.
  100. ## Phase 3: Creator Filtering (Two-Tier)
  101. Filtering is split into **Tier A (Objective Data)** and **Tier B (Subjective Fit)**. Tier A is applied first to narrow the candidate pool, then Tier B is evaluated during deep analysis.
  102. ### Tier A: Objective Data Filtering
  103. #### Priority Order (most important first)
  104. | Priority | Metric | Default Threshold | Description |
  105. |----------|--------|------------------|-------------|
  106. | **P0** | 近期帖子综合质量(1个月内) | 3/5 篇以上"达标"(见下方评估方法) | **最关键指标。** 必须对每篇近期帖子从两个维度独立评估,然后统计达标篇数。不设固定绝对值,而是结合博主自身量级综合判断。 |
  107. | **P1** | Historical hit rate | >= 1 post with 1000+ 赞藏 ever | Proves viral potential exists. Use 赞藏 (likes + saves), not likes alone. |
  108. | **P1** | Comment quality | Comments show genuine interest, not spam/bots | Real engagement vs inflated numbers |
  109. | **P2** | Recent 赞藏 trend | Compare recent posts vs older posts — is 赞藏 growing, stable, or declining? | Declining creators have old viral posts but weak recent numbers. Avoid. |
  110. | **P2** | Total 赞藏/followers ratio | > 3x | Overall engagement health. High ratio from ancient posts is misleading — cross-reference with recent data. |
  111. | **P3** | Follower count | 5,000 - 500,000 (soft reference) | **Soft filter only.** A 3k-follower creator with amazing recent data SHOULD still be recommended. A 50k-follower creator with dead recent data should be REJECTED. |
  112. #### P0 每篇帖子评估方法(两个维度)
  113. 对近期每篇帖子,分别从以下两个维度判断是否健康:
  114. **维度1:赞藏率(赞藏数 / 粉丝数)**
  115. 不设固定百分比,而是看"赞藏数与粉丝量级是否相称":
  116. - 赞藏数远高于粉丝数的一定比例 → 说明内容被算法放大推给了非粉丝 → 健康
  117. - 赞藏数极低,与粉丝量完全不匹配 → 说明内容未被算法认可 → 偏弱
  118. - 横向对比该博主其他帖子:该篇是明显高于还是低于自身均值?
  119. **维度2:赞藏评比例(赞藏数 : 评论数)**
  120. 每篇帖子都需要检查,不只是整体历史数据:
  121. - 20:1 ~ 80:1 → 健康,真实互动
  122. - 80:1 ~ 150:1 → 偏高,结合内容类型判断(教程类收藏高、评论少属正常)
  123. - 150:1 ~ 200:1 → 明显偏高,需留意,建议人工核查评论质量
  124. - > 200:1 → 较大注水嫌疑,该篇视为不达标
  125. **单篇帖子达标判定(结合两个维度):**
  126. - 赞藏率合理 + 赞藏评比例 <150:1 → ✅ 达标
  127. - 赞藏率合理 + 赞藏评比例 150~200:1 → ⚠️ 勉强,酌情处理
  128. - 赞藏率极低 或 赞藏评比例 >200:1 → ❌ 不达标
  129. **P0 总评(根据达标篇数):**
  130. - 5/5 达标 → 优秀
  131. - 4/5 达标 → 良好
  132. - 3/5 达标 → 勉强可用,需在报告中标注
  133. - ≤2/5 → P0 不通过,淘汰
  134. #### Step 1: Profile-Level Quick Screen
  135. For each candidate author, visit their profile page to extract:
  136. - Follower count
  137. - Total likes/favorites
  138. - Bio/description
  139. Use Playwright `browser_run_code` to batch-visit multiple profiles efficiently.
  140. See `guides/platform-selectors.md` for extraction code.
  141. **Soft-reject** candidates far outside follower range (e.g., < 1,000 or > 1,000,000), but keep borderline cases if other signals are strong. Do NOT yet evaluate "recent post quality" from profile-level data.
  142. #### Step 2: Recent Post Data Collection (CRITICAL — Do NOT Skip)
  143. **WARNING: Profile pages do NOT sort notes chronologically.** The notes shown on a profile page are algorithmically ordered and mix old viral posts with recent ones. You CANNOT determine recency from profile page alone.
  144. **最常见的错误:只看到一篇高赞帖就认为达人数据好。必须强制验证多篇近期帖子。**
  145. For each candidate that passes the profile quick-screen, you MUST:
  146. 1. **Click into 5-8 individual posts** from their profile to check actual publish dates
  147. 2. **Identify all posts published within the recency window** (default: 1 month)
  148. 3. **For each verified recent post, record ALL of the following:**
  149. - 发布日期
  150. - 点赞数
  151. - 收藏数
  152. - 赞藏合计 (likes + saves)
  153. - 评论数
  154. - 赞藏评比例 (赞藏 : 评论)
  155. - 赞藏率 (赞藏数 / 粉丝数,百分比)
  156. 4. **对每篇帖子按两个维度独立打标**(参考上方 P0 评估方法):✅ 达标 / ⚠️ 勉强 / ❌ 不达标
  157. 5. **Build a 近期帖子数据表** with at least 3 rows (ideally 5). If a creator has fewer than 3 posts in the last month, they fail P0 automatically.
  158. 6. **统计达标篇数**,得出该创作者的 P0 总评。
  159. **The goal is to present a DATA TABLE for each creator with per-post evaluations. A creator with 4/5 posts consistently healthy is far better than one with 1 viral post and 4 near-zero posts.**
  160. #### Step 3: Apply P0-P3 Filters
  161. Apply filters in priority order. Candidates must pass P0 to proceed. P1-P3 contribute to a composite score but a single P3 failure does not eliminate a candidate with strong P0/P1.
  162. ### Tier B: Subjective Fit Assessment (evaluated in Phase 4)
  163. These dimensions are assessed during deep content analysis. They cannot be automated by data alone -- require reading/watching actual content.
  164. | Dimension | What to Evaluate | How to Check |
  165. |-----------|-----------------|-------------|
  166. | **Content Style Match** | Does the creator's tone match the brand? (专业/轻松/极客/创意) Are they informative, entertaining, or both? | Watch 1-2 videos or read 2-3 posts. Summarize their style in 2-3 words. |
  167. | **AI/Tech Experience** | Has the creator produced AI or tech content before? How deep is their technical understanding? | Scan post history for AI/tech keywords. Check bio for tech credentials. Rate: no experience / surface / moderate / deep. |
  168. | **Competitor Collaboration** | Has the creator worked with competitor products? Which ones? How recently? | Search their post history for competitor product names (from Phase 1). Note: competitor experience can be POSITIVE (proves they understand the space) or NEGATIVE (exclusive deals, brand conflict). |
  169. | **Visual Quality** | Cover design, editing quality, visual consistency, brand-readiness of their content | Screenshot representative posts. Score 1-10 using the Visual Quality Scoring table. |
  170. | **Audience Fit** | Does the creator's audience overlap with the target user base? | Check comment section: are commenters asking relevant questions? Check follower demographics if available. |
  171. | **Brand Safety** | Any controversial content, inappropriate tone, or risky associations? | Quick scan of recent posts for red flags. |
  172. ### Xiaohongshu: Extract Post Date from Detail Page
  173. ```javascript
  174. () => {
  175. // Date is in the note detail page, usually near the bottom of the content area
  176. const dateEl = document.querySelector('[class*="date"], [class*="time"], .bottom-container .date');
  177. const allText = document.body.innerText;
  178. // Match patterns like "02-05", "2025-12-24", "3天前", "昨天"
  179. const dateMatch = allText.match(/(\d{4}-\d{2}-\d{2}|\d{2}-\d{2}|\d+天前|昨天|前天|\d+小时前)/);
  180. return dateMatch?.[0];
  181. }
  182. ```
  183. Alternatively, batch-check by navigating to each post URL and extracting the date element.
  184. ### Output
  185. Present a filtered candidate table with **date-verified** recent post data and search layer source:
  186. ```
  187. | # | Creator | Layer | Followers | Total Likes | Recent Post (date) | Likes | Best Historical | Profile |
  188. ```
  189. Layer column shows: L1 (exact match), L2 (competitor/category), L3 (broader domain)
  190. ## Phase 4: Deep Content Analysis + Subjective Fit (Tier B)
  191. For each creator that passes Tier A filtering, perform deep content analysis AND evaluate all Tier B subjective dimensions simultaneously.
  192. ### 4a. Video Content Analysis
  193. **Bilibili 特殊处理(视频过长,跳过播放):**
  194. 1. **Navigate** to the video page
  195. 2. **Screenshot** the cover image only (do NOT play the video)
  196. 3. **Extract** text info via browser_evaluate:
  197. - Title, description, tags
  198. - Play count, likes, coins, favorites, shares
  199. - Publish date
  200. 4. **Evaluate** based on cover + text only; skip transcription entirely
  201. **其他平台(小红书等)正常流程:**
  202. 1. **Navigate** to the video post
  203. 2. **Play** the video (click video element or play button)
  204. 3. **Screenshot** 3 frames:
  205. - Frame 1: ~2s after start (opening/hook)
  206. - Frame 2: ~middle of video (main content)
  207. - Frame 3: ~near end (conclusion/CTA)
  208. - Save to `screenshots/` directory
  209. 4. **Read** screenshots to evaluate visual quality
  210. 5. **Transcribe** speech if meaningful audio content:
  211. - First check if the note description already contains a transcript (timestamps like "00:00", "00:24")
  212. - If no transcript in text, use the transcribe script:
  213. ```bash
  214. /Library/Frameworks/Python.framework/Versions/3.12/bin/python3 ${CLAUDE_SKILL_DIR}/scripts/transcribe.py "<video_url>"
  215. ```
  216. - Evaluate: clarity, information density, pacing, expertise level
  217. ### 4b. Image/Text Content Analysis
  218. 1. **Navigate** to the post detail page
  219. 2. **Screenshot** the post (cover image + content)
  220. 3. **Extract** full text content via browser_evaluate
  221. 4. **Evaluate**:
  222. - Cover design: brand consistency, color scheme, typography
  223. - Image quality: creativity, clarity, visual impact
  224. - Text quality: information density, actionability, originality
  225. - Title: click-worthiness, emotional hooks
  226. ### 4c. Tier B Subjective Evaluation (do this FOR EVERY candidate)
  227. While analyzing content in 4a/4b, simultaneously evaluate all Tier B dimensions:
  228. #### Content Style Match
  229. - Summarize the creator's tone in 2-3 words (e.g., "专业科普型", "轻松幽默型", "极客实操型")
  230. - Compare against the user's tone preference from Phase 1
  231. - Rate: Mismatch / Partial Match / Strong Match
  232. #### AI/Tech Experience Check
  233. - Scan post history for AI/tech content keywords
  234. - Check bio for tech credentials (company, degree, role)
  235. - Rate the depth: No experience / Surface-level (转载为主) / Moderate (有自己的理解) / Deep (原创深度内容)
  236. #### Competitor Collaboration Scan
  237. - Search the creator's recent posts for competitor product names (from Phase 1 "Known Competitors")
  238. - Record: which competitors, how recently, what type of content (测评/教程/广告)
  239. - Flag: competitor experience is usually POSITIVE (proves they understand the space and can produce brand content), but note any exclusive deals or direct conflicts
  240. #### Audience Fit
  241. - Read 5-10 comments on the creator's best-performing posts
  242. - Are commenters asking relevant questions? Are they potential target users?
  243. - Note any demographic signals (students, professionals, developers, general audience)
  244. #### Brand Safety
  245. - Quick scan recent 10 posts for red flags: controversial topics, inappropriate tone, political content, negative sentiment
  246. - Rate: Safe / Minor concerns / Red flag
  247. ### Visual Quality Scoring (1-10)
  248. | Score | Criteria |
  249. |-------|----------|
  250. | 9-10 | Professional design, strong brand identity, high-res visuals, consistent style |
  251. | 7-8 | Good design sense, clean layout, decent visual impact |
  252. | 5-6 | Acceptable quality, basic design, functional but not distinctive |
  253. | 3-4 | Low effort, screenshots/plain text, no visual identity |
  254. | 1-2 | Poor quality, blurry, cluttered, unprofessional |
  255. ## Phase 5: Final Report
  256. Output a ranked creator report with:
  257. ### Per-Creator Card
  258. ```markdown
  259. ### [Rank]. Creator Name -- Rating: X/5 stars
  260. | Metric | Value |
  261. |--------|-------|
  262. | Platform | Xiaohongshu |
  263. | Search Layer | L1 精准匹配 / L2 竞品赛道 / L3 泛AI领域 |
  264. | Followers | X |
  265. | 总赞藏 (历史累计) | X |
  266. | 历史最高帖 | "title" (赞藏 X) |
  267. | 近期赞藏趋势 | 上升 / 稳定 / 下降 |
  268. | 近期数据稳定性 | 稳定 / 不稳定(最高/最低比 > 5x) |
  269. | Profile | [link](url) |
  270. **近期3-5篇帖子数据(1个月内,必须逐一列出)**
  271. | # | 标题 | 日期 | 赞 | 收藏 | 赞藏 | 评论 | 赞藏评比例 |
  272. |---|------|------|----|------|------|------|-----------|
  273. | 1 | "帖子标题" | MM-DD | X | X | X | X | X:1 |
  274. | 2 | "帖子标题" | MM-DD | X | X | X | X | X:1 |
  275. | 3 | "帖子标题" | MM-DD | X | X | X | X | X:1 |
  276. | 4 | "帖子标题" | MM-DD | X | X | X | X | X:1 |
  277. | 5 | "帖子标题" | MM-DD | X | X | X | X | X:1 |
  278. 近期均值:赞藏 X,评论 X,赞藏评比例 X:1
  279. **Tier A 数据评分**
  280. - 近期赞藏一致性 (P0): X/10 — [描述:如"5篇均超2000赞藏,最高/最低比仅1.3x,极稳定"]
  281. - 历史爆款 (P1): X/10
  282. - 赞藏评比例 (P1): X/10 — 赞藏:评论 = X:1(正常范围 20-80:1,>100:1 有注水风险)
  283. - 赞藏趋势 (P2): X/10
  284. **Tier B 调性评分**
  285. - 内容风格: [2-3词描述] — [匹配度: Strong/Partial/Mismatch]
  286. - AI/科技经验: [No/Surface/Moderate/Deep] — [具体说明]
  287. - 竞品合作: [有/无] — [具体产品及时间]
  288. - 视觉质量: X/10
  289. - 受众匹配: [描述目标受众重合度]
  290. - 品牌安全: [Safe/Minor concerns/Red flag]
  291. **综合评价**: [2-3 sentence summary covering: why this creator fits, what kind of content they could produce for the brand, any risks or considerations]
  292. **Representative Screenshots**: [reference saved screenshot files]
  293. ```
  294. ### Summary Table
  295. ```
  296. | Rank | Creator | Layer | Followers | 近期均赞藏 | 历史最高赞藏 | Style | AI Exp | Competitor | Visual | Overall |
  297. ```
  298. ### Ranking Logic
  299. Final ranking considers both Tier A and Tier B, weighted as follows:
  300. - **50%** Tier A (objective data): recent engagement is dominant factor
  301. - **30%** Tier B relevance: content style match + AI experience + competitor knowledge
  302. - **20%** Tier B quality: visual quality + audience fit + brand safety
  303. A creator with average data but perfect brand fit may rank higher than one with great data but poor style match. Flag any cases where Tier A and Tier B scores diverge significantly -- let the user make the final call.
  304. ## Dependencies
  305. ### Required (check on first run)
  306. ```bash
  307. # Check and prompt if missing:
  308. which ffmpeg || echo "MISSING: Run 'brew install ffmpeg'"
  309. python3 -c "import whisper" 2>/dev/null || echo "MISSING: Run 'pip install openai-whisper'"
  310. which opencli || echo "MISSING: opencli not installed (use Playwright fallback)"
  311. ```
  312. ### Screenshots Directory
  313. Create `screenshots/` in the current working directory for storing captured frames.
  314. ## High-Potential Small Creator Profile (Reference Model)
  315. When evaluating small-but-promising creators, look for these traits. Based on real case: 浙大猫学长 (1.2万粉, 9.1万获赞, 18篇帖子, 多篇近期1000+赞).
  316. ### Key Indicators
  317. | Trait | Why It Matters | Example |
  318. |-------|---------------|---------|
  319. | **Low post count, high hit rate** | Quality > quantity. A creator with 20 posts and 7 viral hits (35%+ hit rate) is better than one with 200 posts and 3 hits. | 18 posts, 7 posts with 1000+ likes = 39% hit rate |
  320. | **Extreme engagement ratio** | Best post likes / follower count > 50% signals strong content-market fit and algorithmic recommendation. | 9126 / 12000 = 76% |
  321. | **Real professional credentials** | Domain expertise in bio (specific company, degree, role) that matches content topic. Not vague "AI enthusiast". | "浙大数学硕士 \| 腾讯AI算法工程师" |
  322. | **Niche pioneer** | Creating or defining a new category rather than following trends. First-mover content gets disproportionate algorithmic boost. | Coined "Vibe Animation" / "Vibe Motion" as new categories |
  323. | **Consistent recent high performance** | Multiple 1000+ posts in last 6 weeks, not just one lucky viral hit. | 4 posts with 1000+ likes in 6 weeks |
  324. | **Video-first with technical depth** | Video content with real demos/tutorials, not just talking heads or text overlays. Shows actual technical ability. | Live coding demos, animation tutorials |
  325. | **Getting ratio > 5x** | Total likes-and-saves / followers > 5x means content reaches far beyond existing audience. | 9.1万 / 1.2万 = 7.6x |
  326. ### Red Flags (Anti-Patterns)
  327. | Red Flag | What It Signals |
  328. |----------|----------------|
  329. | High followers but low recent likes | Declining creator, bought followers, or topic fatigue |
  330. | Viral posts are all 3+ months old | One-hit wonder or abandoned account |
  331. | Many posts with < 50 likes mixed with rare 1000+ | Inconsistent quality, likely gaming the algorithm |
  332. | Bio says "AI博主" but content is generic reposts | No original perspective, low defensibility |
  333. | Post titles are clickbait but content is thin | Short-term tactic, audience will churn |
  334. | Very high posting frequency (daily) with low avg likes | Volume strategy without quality, unlikely to produce breakthroughs |
  335. ## Common Pitfalls (Lessons Learned)
  336. ### 1. Profile page notes are NOT chronologically sorted
  337. The notes displayed on a Xiaohongshu profile page are algorithmically ranked, NOT by date. A profile might show a 6-month-old viral post (9000 likes) next to a 2-day-old post (30 likes). **Never assume profile-visible posts are "recent".**
  338. ### 2. Must click into posts to verify dates
  339. The profile page extraction only gives you titles and like counts -- no dates. You MUST navigate into individual post detail pages to see the actual publish date. Without this step, you will mistake old viral posts for current performance.
  340. ### 3. High historical likes != current quality
  341. A creator with a 10,000-like post from 6 months ago might now only get 50-200 likes per post. This indicates declining engagement. Always compare **date-verified recent performance** against the user's threshold, not historical peaks.
  342. ### 4. Search result date hints
  343. In Xiaohongshu search results, the `author` field extracted via JS often includes a date suffix (e.g., "博主名02-05", "博主名2天前"). Parse these as a preliminary recency signal before investing time in profile visits.
  344. ### 5. Batch profile visits save time, but don't skip date verification
  345. It's efficient to batch-visit profiles for follower counts (quick-reject on follower threshold). But after the follower screen, you MUST do a second pass clicking into actual posts. Don't skip this step to save time -- it leads to recommending declining creators.
  346. ### 6. Total likes/followers ratio can be misleading
  347. A very high ratio (e.g., 26x) might come from a few ancient viral posts rather than consistent recent performance. Always cross-reference with date-verified recent post data.
  348. ### 7. Detail page like counts extraction is unreliable
  349. The JS selectors for extracting likes from a post detail page often grab wrong elements (comment counts, "赞" button text, numbers from titles). **Use the profile-level likes data** (which is more reliable) for actual like counts, and use detail page visits ONLY for date verification. Cross-reference: profile gives you (title, likes), detail page gives you (title, date) -- match them by title.
  350. ## Notes
  351. - Always use opencli first; fall back to Playwright only if opencli fails or doesn't support the platform
  352. - When using Playwright, close login modals before extracting data
  353. - Video transcription is optional -- skip if the note text already contains a transcript
  354. - Keep screenshot count reasonable (3 per video, 1-2 per image post) to save context
  355. - Present results in Chinese (the user's preferred language)
  356. - Rate creators relative to each other, not on absolute scales
  357. ## Exporting to Feishu Spreadsheet
  358. When user asks to export results to a Feishu spreadsheet, follow `guides/feishu-export.md`.
  359. **Key rule**: Always use `lark_api` PUT directly — never use `lark_sheet` write action.
  360. `lark_sheet` write silently fails (returns `{"data": {}}` with no error) and data won't appear in the sheet.
  361. Quick template:
  362. ```
  363. 1. create_spreadsheet(title="...") → get spreadsheet_token
  364. 2. get_spreadsheet_sheets(token) → get sheetId
  365. 3. lark_api PUT /open-apis/sheets/v2/spreadsheets/{token}/values
  366. body: { "valueRange": { "range": "{sheetId}!A1:H{n}", "values": [[...], ...] } }
  367. 4. Verify response has updatedRows == expected count
  368. ```
README.md · README · 109 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. <div align="center">
  2. # 🔍 find-influencer Skill
  3. **一个让 AI Agent 帮你跨平台找达人的 Skill**
  4. 支持小红书、抖音、B 站、YouTube 等多平台。从"我想找 XXX 类型的博主"到"完整的达人评估报告",全流程自动化。
  5. [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
  6. [![Skill: Claude Code](https://img.shields.io/badge/Skill-Claude%20Code-8A2BE2)](https://github.com/anthropics/claude-code)
  7. [![Platforms](https://img.shields.io/badge/Platforms-小红书%20·%20抖音%20·%20B站%20·%20YouTube-FF2741)]()
  8. [![Stars](https://img.shields.io/github/stars/cool111111/find-influencer-skill?style=social)](https://github.com/cool111111/find-influencer-skill/stargazers)
  9. </div>
  10. > 做内容运营的人都知道:找达人这件事 80% 的时间花在了**翻平台、看主页、评估内容质量**上。
  11. > 这个 Skill 把这 80% 自动化了。
  12. <!-- TODO 在此处插入一张 demo GIF 或截图(达人发现 + 评估报告产出过程) -->
  13. ## 解决什么问题
  14. 传统做法:
  15. - 打开小红书,搜关键词,一个个点开主页看
  16. - 翻 50 个主页 + 看 200 条内容,挑出 5 个值得对接的
  17. - 整理成 Excel,写评估笔记
  18. - **耗时:3-5 小时**
  19. 用这个 Skill:
  20. - 描述需求:"**帮我找 10 个小红书上做 AI 教程的中腰部博主**"
  21. - Agent 自动跨平台搜索 → 数据筛选 → 深度内容分析 → 输出结构化评估报告
  22. - **耗时:15-20 分钟**
  23. ## 工作流
  24. ```
  25. 1. 收集需求(领域 / 平台 / 粉丝量级 / 内容类型 / 商业化偏好)
  26. ↓
  27. 2. 跨平台搜索(opencli / Playwright 自动化)
  28. ↓
  29. 3. 数据筛选(粉丝量 / 互动率 / 更新频率)
  30. ↓
  31. 4. 深度内容分析(抽样视频转写 + 内容质量评估)
  32. ↓
  33. 5. 结构化报告(达人画像 + 优劣势 + 合作建议)
  34. ```
  35. ## Quick Start
  36. ```bash
  37. # 1. Clone Skill
  38. git clone https://github.com/cool111111/find-influencer-skill.git \
  39. ~/.your-agent/skills/find-influencer
  40. # 2. Check dependencies
  41. bash ~/.your-agent/skills/find-influencer/scripts/setup_check.sh
  42. # 3. 在 Agent 里说:
  43. # "帮我找 10 个小红书上做 XXX 的博主"
  44. ```
  45. ## 依赖项
  46. ### 必需
  47. | 依赖 | 用途 | 安装 |
  48. |------|------|------|
  49. | Python 3 | 视频转写脚本 | `brew install python` / `apt install python3` |
  50. | ffmpeg | 视频音频抽取 | `brew install ffmpeg` / `apt install ffmpeg` |
  51. | openai-whisper | 语音转文字 | `pip install openai-whisper` |
  52. ### 可选
  53. | 依赖 | 用途 | 兜底方案 |
  54. |------|------|---------|
  55. | opencli | CLI 搜索小红书 / B 站 / YouTube | Playwright 浏览器自动化 |
  56. ### Agent 环境
  57. | 要求 | 用途 |
  58. |------|------|
  59. | Claude Code(或兼容的 Skill runtime) | Skill 运行环境 |
  60. | Playwright MCP Server | 抖音 / 主页访问 / 内容分析的浏览器自动化 |
  61. ## 一键安装(macOS)
  62. ```bash
  63. brew install python ffmpeg
  64. pip install openai-whisper
  65. ```
  66. ## 实战数据
  67. 这个 Skill 在实际达人运营场景中验证过,覆盖大量创作者的初筛 + 评估。实测反馈:从"找达人 → 输出对接清单"的总时长压缩了 **80%+**。
  68. ## 注意事项
  69. - **首次运行会下载 Whisper `base` 模型**(约 140MB)
  70. - **平台访问需要登录态** — 小红书、B 站等需要在 Chrome 提前登录,opencli/Playwright 会复用浏览器登录态
  71. - **转写自动跳过** — 如果帖子已有文字稿,会跳过视频转写步骤
  72. ## 配套 Skill
  73. - 长文写作 → [Writer Skill](https://github.com/cool111111/writer-skill)
  74. - 小红书短图文 → [XHS Content Skill](https://github.com/cool111111/xhs-content-skill)
  75. ## License
  76. MIT — 觉得有用就 **star 一下** ⭐

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.