小红书 Research 📕

Research tool for Chinese user-generated content — travel, food, lifestyle, local discoveries.

When to Use

• Travel planning and itineraries
• Restaurant/cafe/bar recommendations
• Activity and weekend planning
• Product reviews and comparisons
• Local discovery and hidden gems
• Any question where Chinese perspectives help

Recommended Model

When spawning as a sub-agent: Sonnet 4.5 (model: "claude-sonnet-4-5-20250929")

• Fast enough for the slow XHS API calls
• Good at Chinese content understanding
• More cost-effective than Opus for research grunt work
• Opus overkill for search → synthesize workflow

Context Management (Always Use)

ALWAYS use dynamic context monitoring — even 5 posts with images can hit 75-300k tokens.

The Problem

• Each post with images = 15-60k tokens
• 200k context fills fast
• Context is append-only (can’t “forget” within session)

The Solution: Monitor + Checkpoint + Continue

1. After EACH post, do two things:

a) Write findings to disk immediately:
   /research/{task-id}/findings/post-{n}.md

b) Check context usage:
   session_status → look for "Context: XXXk/200k (YY%)"

2. When context hits 70%, STOP and checkpoint:

Write state file:
/research/{task-id}/state.json
{
  "processed": 15,
  "pendingUrls": ["url16", "url17", ...],
  "summaries": ["Post 1: 火塘...", ...]
}

Return to caller:
{
  "complete": false,
  "processed": 15,
  "remaining": 25,
  "statePath": "/research/{task-id}/state.json",
  "findingsDir": "/research/{task-id}/findings/"
}

3. Caller spawns fresh sub-agent to continue:

spawn_subagent(
  task="Continue XHS research from /research/{task-id}/state.json",
  model="claude-sonnet-4-5-20250929"
)

New sub-agent has fresh 200k context, reads state.json, continues from post 16.

State File Schema

{
  "taskId": "kunming-food-2026-02-01",
  "query": "昆明美食",
  "searchesCompleted": ["昆明美食", "昆明美食推荐"],  // Keywords already searched
  "processedUrls": ["url1", "url2", ...],             // Explicit URL tracking (prevents duplicates)
  "pendingUrls": ["url3", "url4", ...],               // Remaining URLs to process
  "nextPostNumber": 16,                                // Next post-XXX.md number
  "summaries": [                                       // 1-liner per post for final synthesis
    "Post 1: 火塘餐厅 | 🟢 | ¥80 | 本地人推荐",
    "Post 2: 野生菌火锅 | 🟢 | ¥120 | 菌子新鲜"
  ],
  "batchNumber": 1,
  "contextCheckpoint": "70%"
}

Critical fields for handoff:

• processedUrls: Prevents re-processing same post across sub-agents
• pendingUrls: Exact work remaining
• nextPostNumber: Ensures sequential file naming
• searchesCompleted: Prevents duplicate searches

Workflow for Large Research

Caller should use longer timeout:

sessions_spawn(
  task="...",
  model="claude-sonnet-4-5-20250929",
  runTimeoutSeconds=1800  // 30 minutes for research tasks
)

Default is 600s (10 min) — too short for XHS research with slow API calls.

Interleave search and processing (don’t collect all URLs first):

[XHS Sub-agent 1]
    ├── Check for state.json (none = fresh start)
    ├── Search keyword 1 → get 20 URLs
    ├── Process 5-10 posts immediately (writing each to disk)
    ├── Search keyword 2 → get more URLs (dedupe)
    ├── Process more posts
    ├── Context hits 70% → write state.json
    └── Return {complete: false, remaining: N}

This prevents timeout from losing all work — each post is saved as processed.

Full continuation pattern:

[Caller]
    ↓ spawn (runTimeoutSeconds=1800)
[XHS Sub-agent 1]
    ├── Search + process interleaved
    ├── Context hits 70% → write state.json
    └── Return {complete: false, remaining: 25}
    
[Caller sees incomplete]
    ↓ spawn continuation (runTimeoutSeconds=1800)
[XHS Sub-agent 2]  ← fresh 200k context!
    ├── Read state.json (has processedUrls, pendingUrls)
    ├── Continue processing + more searches if needed
    ├── Context hits 70% → write state.json
    └── Return {complete: false, remaining: 10}
    
[Caller sees incomplete]
    ↓ spawn continuation
[XHS Sub-agent 3]
    ├── Read state.json
    ├── Process remaining posts
    ├── All done → write synthesis.md
    └── Return {complete: true, synthesisPath: "..."}

Output Directory Structure

/research/{task-id}/
├── state.json              # Checkpoint for continuation
├── findings/
│   ├── post-001.md         # Full analysis + image paths
│   ├── post-002.md
│   └── ...
├── images/
│   ├── post-001/
│   │   ├── 1.jpg
│   │   └── 2.jpg
│   └── ...
├── summaries.md            # All 1-liners (for quick scan)
└── synthesis.md            # Final output (when complete)

Key Rules (ALWAYS FOLLOW)

1. Write after EVERY post — crash-safe, no work lost
2. Check context after EVERY post — use session_status tool
3. Stop at 70% — leave room for synthesis + buffer
4. Return structured result — caller decides next step
5. Read all images — they’re pre-compressed (600px, q85)
6. Skip videos — already marked in fetch-post

⚠️ This is not optional. Even small research can overflow context with image-heavy posts.

Scripts (Mechanical Tasks)

These scripts handle the repetitive CLI work:

All scripts are at /root/clawd/skills/xhs/bin/

Preflight (always run first)

/root/clawd/skills/xhs/bin/preflight

Checks: rednote-mcp installed, cookies valid, stealth patches, test search. Don’t proceed until preflight passes.

Search

/root/clawd/skills/xhs/bin/search "昆明美食推荐" [limit] [timeout] [sort]

Returns JSON with post results.

Parameters:

Sort options:

Examples:

# Default sort (recommended for most research)
bin/search "昆明美食推荐" 20

# Recent posts first (舆情, current events)
bin/search "某品牌 评价" 20 180 newest

# Most popular posts
bin/search "网红打卡地" 15 180 hot

Scroll pagination enabled (patched): When limit > 20, the tool scrolls to load more results via XHS infinite scroll. Actual results depend on available content.

For maximum coverage, combine:

1. Higher limits (e.g., limit=50) to scroll for more
2. Multiple keyword variations for different result sets:
   • 香蕉攀岩, 香蕉攀岩馆, 香蕉攀岩体验, 香蕉攀岩评价
   • 昆明美食, 昆明美食推荐, 昆明必吃, 昆明本地人推荐

Results vary by query — popular topics may return 30-50+, niche topics fewer.

Choosing sort order:

• Most research → general (default). Let XHS’s algorithm surface the best content.
• 舆情监控 / sentiment tracking → newest. You want recent opinions, not old viral posts.
• Trend discovery → hot. See what’s currently popular.

Get Content

/root/clawd/skills/xhs/bin/get-content "FULL_URL_WITH_XSEC_TOKEN"

⚠️ Must use full URL with xsec_token from search results.

Get Comments

/root/clawd/skills/xhs/bin/get-comments "FULL_URL_WITH_XSEC_TOKEN"

Get Images

Download all images from a post to local files:

/root/clawd/skills/xhs/bin/get-images "FULL_URL" /tmp/my-images

Fetch Post (Deep Dive with Images)

Fetch content, comments, and images in one call — with built-in retries:

/root/clawd/skills/xhs/bin/fetch-post "FULL_URL" /path/to/cache [max_retries]

Features:

• Retries on timeout (60s → 90s → 120s)
• Clear error reporting in JSON output
• Images cached locally, bypassing CDN protection

Returns JSON:

{
  "success": true,
  "postId": "abc123",
  "content": { 
    "title": "...", 
    "author": "...", 
    "desc": "...", 
    "likes": "983", 
    "tags": [...],
    "postDate": "2025-09-04"  // ← Added via patch!
  },
  "comments": [{ "author": "...", "content": "...", "likes": "3" }, ...],
  "imagePaths": ["/cache/images/abc123/1.jpg", ...],
  "errors": []
}

Date filtering: Use postDate to filter out old posts. Skip posts older than your threshold (e.g., 6-12 months for restaurants).

Workflow:

1. fetch-post → JSON + cached images
2. Read each imagePath directly (Claude sees images natively)
3. Combine text + comments + what you see into findings

Viewing images:

Read("/path/to/1.jpg")  # Claude sees it directly - no special tool needed

Look for: visible text (addresses, prices, hours), atmosphere, food presentation, crowd levels.

Research Methodology (Judgment Tasks)

This is where you think. Scripts do the fetching; you do the analyzing.

Depth Levels

Minimum is 5 — unless fewer exist. Note limited coverage if <5 results.

Research Workflow

Step 0: Preflight

Run bin/preflight. Don’t proceed until it passes.

Step 1: Plan Your Searches

Think: “What would a Chinese user search on 小红书?”

• Include location when relevant
• Add qualifiers: 推荐, 攻略, 测评, 探店, 打卡, 避坑
• Consider synonyms and variations
• Plan 2-3 different search angles

Date filtering: Posts include postDate field (e.g., “2025-09-04”). The calling agent specifies the date filter based on research type:

Caller should specify in task description, e.g.:

• “Only posts from last 30 days” (舆情)
• “Posts from 2025 or later” (travel)
• “No date filter” (general research)

If no filter specified: Default to 12 months (safe middle ground).

Fallback when postDate is null: Use keyword hints: 2025, 最近, 最新

Language strategy:

Step 2: Search & Scan

Run your searches. Results are already ranked by XHS’s algorithm (relevance + engagement).

Use judgment based on preview — like a human deciding what to click:

Think: “Given my research goal, would this post likely contain useful information?”

Key insight: For 舆情监控, a 3-like complaint post may be more valuable than a 500-like promotional post. Engagement ≠ relevance for all research types.

Step 3: Deep Dive Each Post

For each selected post, use fetch-post to get everything:

bin/fetch-post "url_from_search" {{RESEARCH_DIR}}/xhs

Returns JSON with content, comments, and cached images. Has built-in retries. Then:

A. Review content

• Extract key facts from title/description
• Note author’s perspective/bias
• Check tags for categorization

B. View images (critical!) For each imagePath in the result, just read it:

Read("/path/to/1.jpg")  # You see it directly

• Look for text overlays: addresses, prices, hours
• Note visual details: ambiance, crowd levels, food presentation

⚠️ Don’t describe images in isolation. Synthesize what you see with the post content and comments to form a holistic view. An image of a crowded restaurant + author saying “周末排队1小时” + comments confirming “人超多” = that’s your finding about crowds.

C. Review comments (gold for updates)

• “已经关门了” = already closed
• Real experiences vs sponsored hype
• Tips not in main post

D. Return picked images Include paths to the best/most informative images in your findings. The calling agent decides whether and how to use them (embed in reports, reference, etc.). You’re curating — pick images that show something useful (venue exterior, menu with prices, actual food, atmosphere) not just decorative shots.

Step 4: Synthesize

• What do multiple sources agree on?
• Any contradictions?
• What’s the overall consensus?
• What would you actually recommend?

Step 5: Output

Facts + Flavor — structured findings that preserve the XHS voice.

## XHS Research: [Topic]

### Search Summary
| Search | Results | Notes |
|--------|---------|-------|
| 昆明攀岩 | 10 | Good coverage |

### Findings

#### [Venue Name] (中文名)
- **Type:** Restaurant / Activity / Attraction
- **Address:** [from post or image]
- **Price:** ¥XX/person
- **Hours:** [if found]
- **The vibe:** [atmosphere, energy — preserved voice]
- **Why people like it:** [opinions, impressions]
- **Watch out for:** [warnings from comments]
- **Source:** [full URL]
- **Engagement:** X likes
- **Images:** [paths for calling agent to use]
  - `/path/to/1.jpg` — exterior/entrance
  - `/path/to/3.jpg` — menu with prices

> "引用原文..." — @username

### Overall Impressions
- Consensus across posts
- Patterns in preferences
- Things only locals know
- Disagreements worth noting

The XHS value is the human perspective. A recommendation that says “环境一般但是味道绝了” tells you more than “Rating: 4.2/5”.

Think: “What would a friend who just spent an hour on XHS tell me?”

Quality Signals

Trustworthy:

• 100+ likes with real comments
• Detailed personal experience
• Multiple photos from actual visit
• Specific details (prices, hours)
• Recent posts (look for date mentions in content: “上周”, “昨天”, “2025å¹´X月”)
• Year in title (e.g., “2025上海咖啡必喝榜”)

Checking recency:

• Look for dates in post text/title
• Check if prices seem current
• Comments mentioning “还在吗” or “现在还有吗” = might be outdated
• Comments with recent dates confirm post is still relevant

Suspicious:

• 广告/合作/赞助 markers
• Overly positive, no specifics
• Stock photos only
• No comments or generic ones
• Very old posts

Timing & Efficiency

XHS is SLOW — Plan Accordingly

The rednote-mcp CLI is slow (30-90s per search). Don’t rapid-fire poll.

When running searches via exec:

# GOOD: Give it time to complete
exec(command, yieldMs: 60000)  # Wait 60s before checking
process(poll)  # Then poll every 30s if still running

DON’T:

• Poll every 2-3 seconds (wastes tokens, no benefit)
• Start multiple searches simultaneously (overloads)
• Wait indefinitely without writing partial results

Write Incrementally

Don’t wait until you’ve analyzed everything to start writing. After each batch of 3-5 posts:

• Append findings to your output file
• This protects against timeout/termination losing all work

## Findings (in progress)

### Batch 1: 美食搜索 (3 posts analyzed)
[findings...]

### Batch 2: 攻略搜索 (analyzing...)

Time Budget Awareness

If you’ve been running 15+ minutes:

• Prioritize writing what you have
• Note incomplete searches in output
• Better to deliver 80% findings than lose 100% to termination

Retry Pattern

rednote-mcp is slow. If a command times out:

Attempt 1: default timeout
Attempt 2: +60s
Attempt 3: +120s

If all fail, report the failure. Do NOT fall back to web_search — defeats the purpose.

Error Handling

Setup & Maintenance

First-Time Setup

npm install -g rednote-mcp
npx playwright install
/root/clawd/skills/xhs/patches/apply-all.sh
xvfb-run -a rednote-mcp init

Re-login (when cookies expire)

xvfb-run -a rednote-mcp init

After rednote-mcp updates

/root/clawd/skills/xhs/patches/apply-all.sh

Role Clarification

This skill = Research tool that outputs structured findings Calling agent = Synthesizes XHS + other sources into final reports, decides which images to embed

You return:

• Synthesized findings (text + images + comments → holistic view)
• Curated image paths (calling agent decides how to use them)
• Preserved human voice (opinions, vibes, tips)

You don’t:

• Describe images in isolation (“I see a restaurant…”)
• Generate final reports (that’s the caller’s job)
• Decide image layout/placement

XHS is like having a Chinese-speaking friend spend an hour researching for you. They’d give you facts, but also opinions, vibes, and insider tips. That’s what you’re capturing.

Remember: Research like a curious human. Explore, cross-reference, look at pictures, read comments. The “这家真的绝了” matters as much as the address.