Creator Discovery & Analysis

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).

ARGUMENTS: User's initial request (platform, content direction, etc.)

Phase 1: Requirements Gathering

Use AskUserQuestion to collect:

1. Target Platform(s): Xiaohongshu / Douyin / Bilibili / YouTube / TikTok / X(Twitter) / Instagram / multiple
2. Product/Brand: What product or brand is this collaboration for? e.g., ChatGPT, Claude, Gemini
3. Content Direction: The specific topic AND broader category. e.g., "ProductX (broader: AI agent, AI工具)"
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
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
6. Follower Range: default 5,000 - 500,000 (soft reference, not hard cutoff -- great recent data can override low followers)
7. Recent Data Priority: default 1 month. Minimum acceptable recent post engagement (likes/saves/views)
8. Known Competitors: List competitor products so we can check if creators have collaborated with them. e.g., Coze, Dify, FastGPT
9. Number of Results: default 5-10 creators

If ARGUMENTS already contain these details, skip redundant questions.

Phase 2: Content Search

Tool Selection

Search Strategy: Multi-Layer Keyword Expansion

Don't just search the exact product name -- use a 3-layer keyword strategy to find both vertical and adjacent creators:

Layer 1: Exact product/brand keywords (find creators already covering your product)

• Direct product name and variations
• Example for ProductX: "ProductX", "productx教程", "productx测评"

Layer 2: Competitor & category keywords (find creators in the same space)

• Competitor product names from Phase 1
• Category terms that your product belongs to
• Example: "Coze教程", "AI agent平台", "AI工作流搭建", "Dify教程"

Layer 3: Broader domain keywords (find quality AI creators who could pivot to your product)

• Broader topic keywords in the same domain
• Adjacent content areas where your audience overlaps
• Example: "AI工具推荐", "AI效率提升", "AI科技测评", "AI产品体验"

Execution:

1. Generate 2-3 keywords per layer (6-9 total), search each with sort=popularity_descending when available
2. Extract: title, author name, author ID/URL, likes count, saves count (收藏), post date
3. Tag each result with its source layer -- Layer 1 hits are highest relevance, Layer 3 are expansion candidates
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.
5. Filter to recent posts (within 1 month by default) with 赞藏数 (likes + saves) meeting user's threshold
6. Deduplicate by author -- prefer candidates with MULTIPLE recent high-engagement posts over one-hit wonders
7. When deduplicating, preserve the highest-relevance layer tag (if a creator appears in both Layer 1 and Layer 3, tag as Layer 1)

Platform-Specific Search

Xiaohongshu via opencli:

opencli xiaohongshu search "<keyword>" --limit 20 -f json

If blocked by login wall, use Playwright:

Navigate: https://www.xiaohongshu.com/search_result?keyword=<encoded>&type=1&sort=popularity_descending
Close login modal if present, then extract via browser_evaluate

Bilibili via opencli:

opencli bilibili search --keyword "<keyword>" --limit 20 -f json

TikTok via opencli:

opencli tiktok search "<keyword>" --limit 20 -f json
# Get profile stats
opencli tiktok profile <username> -f json
# Get recent videos
opencli tiktok user <username> --limit 20 -f json

X (Twitter) via opencli:

opencli twitter search "<keyword>" --limit 20 -f json
# Get profile stats (followers, bio)
opencli twitter profile <username> -f json

Instagram via opencli:

opencli instagram search "<keyword>" --limit 20 -f json
# Get profile stats
opencli instagram profile <username> -f json
# Get recent posts
opencli instagram user <username> --limit 20 -f json

YouTube via opencli:

opencli youtube search --query "<keyword>" --limit 20 -f json
# Get video metadata (views, likes)
opencli youtube video "<url>" -f json

For YouTube channel subscriber count, use Playwright to visit https://www.youtube.com/@<handle>.

Refer to guides/platform-selectors.md for DOM selectors and JS extraction code.

Phase 3: Creator Filtering (Two-Tier)

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.

Tier A: Objective Data Filtering

Priority Order (most important first)

P0 每篇帖子评估方法（两个维度）

对近期每篇帖子，分别从以下两个维度判断是否健康：

维度1：赞藏率（赞藏数 / 粉丝数）

不设固定百分比，而是看"赞藏数与粉丝量级是否相称"：

• 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
• 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
• 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？

维度2：赞藏评比例（赞藏数 : 评论数）

每篇帖子都需要检查，不只是整体历史数据：

• 20:1 ～ 80:1 → 健康，真实互动
• 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）
• 150:1 ～ 200:1 → 明显偏高，需留意，建议人工核查评论质量

• 200:1 → 较大注水嫌疑，该篇视为不达标

单篇帖子达标判定（结合两个维度）：

• 赞藏率合理 + 赞藏评比例 <150:1 → ✅ 达标
• 赞藏率合理 + 赞藏评比例 150~200:1 → ⚠️ 勉强，酌情处理
• 赞藏率极低 或 赞藏评比例 >200:1 → ❌ 不达标

P0 总评（根据达标篇数）：

• 5/5 达标 → 优秀
• 4/5 达标 → 良好
• 3/5 达标 → 勉强可用，需在报告中标注
• ≤2/5 → P0 不通过，淘汰

Step 1: Profile-Level Quick Screen

For each candidate author, visit their profile page to extract:

• Follower count
• Total likes/favorites
• Bio/description

Use Playwright browser_run_code to batch-visit multiple profiles efficiently. See guides/platform-selectors.md for extraction code.

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.

Step 2: Recent Post Data Collection (CRITICAL — Do NOT Skip)

WARNING: Profile pages do NOT sort notes chronologically. The notes shown on a profile page are algorithmically ordered and mix old viral posts