层 vault 位置 内容 入口 记忆层 workspace（本 vault）~/.openclaw/workspace日档 memory/、长期记忆MEMORY.md、行为规范AGENTS.md、反思DREAMS.md[[memory/README]] 知识库 wiki~/.openclaw/workspace/wiki结论、案例、SOP —— 可复用产出 [[wiki/README]] 系统层 openclaw~/.openclaw技能、扩展、配置 —— 运行底座 该 vault 的 README 读什么去哪：查结论/案例 →
wikivault；查"当时怎么想的" → 本 vaultmemory/；排障改配置 →openclawvault。详细分工与目录结构见
wiki/README.md与memory/README.md。
Scenic-Area-Marketing-CN is a production-grade, fully automated scenic marketing intelligence system built on the OpenClaw AI Agent framework. It runs 24/7 on a Mac Mini, autonomously collecting data from 8+ platforms, generating structured competitive intelligence reports, pushing decision-grade insights to Feishu (飞书) group chats, and archiving everything into an Obsidian knowledge graph.
The system serves Jianye Movie Town (建业电影小镇), a cultural-tourism scenic spot in Zhengzhou, Henan, China, with an annual visitor target of 1.53 million and revenue target of ¥120 million for 2026.
This is not a demo or prototype. It has been running continuously for 30+ days without human intervention, producing 6-8 actionable decision briefs daily.
┌─────────────────────────────────────────────────────────────────────────────────────┐
│ 🧠 AI AGENT CORE │
│ OpenClaw Runtime · DeepSeek-V4-Flash · 23 Cron Tasks │
│ Web Search · Memory · CDP Browser · Feishu API │
└─────────────────────────────────────────────────────────────────────────────────────┘
│ │ │
▼ ▼ ▼
┌─────────────────────┐ ┌─────────────────────────────┐ ┌─────────────────────────┐
│ 📊 DATA LAYER │ │ 🧠 INTELLIGENCE LAYER │ │ ⚡ EXECUTION LAYER │
├─────────────────────┤ ├─────────────────────────────┤ ├─────────────────────────┤
│ Douyin Index Tracker │ │ Decision Framework Engine │ │ Feishu Card Pusher │
│ Xiaohongshu Monitor │ │ Competitive Intel Analysis │ │ (send_feishu_card.py) │
│ Weibo Hot Search │ │ Trend & Anomaly Detection │ │ Cron Job Scheduler │
│ Baidu Search │ │ Viral Content Deconstruction │ │ Cookie Rotation Manager │
│ Passenger CSV Reader │ │ Weekly Pattern Forecasting │ │ CDP Browser Control │
│ CDP Cookie Manager │ │ Learning Loop & Validation │ │ Health Monitor │
└─────────────────────┘ └─────────────────────────────┘ └─────────────────────────┘
│ │ │
└─────────────────────────────┼───────────────────────────┘
▼
┌─────────────────────────────┐
│ 📚 KNOWLEDGE LAYER │
├─────────────────────────────┤
│ Obsidian Wiki (3-Layer Abstraction) │
│ → Concepts / Entities / Sources │
│ → 20 Competitor Profiles │
│ → 10 Viral Formulas / 22 SOPs │
│ → Historical Data (2023-2026) │
└─────────────────────────────┘
┌──────────┐ ┌──────────────┐ ┌───────────────┐ ┌────────────┐
│ 08:00 │───▶│ 10:00-10:30 │───▶│ 13:00-16:00 │───▶│ 18:00-22:00│
│ Cookie │ │ Douyin + XHS │ │ Intel + Hit │ │ Review + │
│ Sync │ │ Daily Report │ │ Deep Analysis │ │ Archive │
└──────────┘ └──────────────┘ └───────────────┘ └────────────┘
│ │ │ │
▼ ▼ ▼ ▼
┌──────────┐ ┌──────────────┐ ┌───────────────┐ ┌────────────┐
│ 3 Cookie │ │ Parsed Data +│ │ Decision │ │ Memory + │
│ Files │ │ Decision │ │ Briefs (D1-D4)│ │ Wiki + │
│ (JuLiang, │ │ Layer │ │ + Risk Alert │ │ GitHub │
│ XHS, Weibo)│ │ Output │ │ │ │ Commit │
└──────────┘ └──────────────┘ └───────────────┘ └────────────┘
│ │
▼ ▼
┌──────────────────────────────┐
│ Feishu Group Chat │
│ (oc_2581c03b79e4893cc36...) │
│ Interactive Card via │
│ send_feishu_card.py │
└──────────────────────────────┘
1. Cron triggers isolated Agent session
│
2. Agent reads task description & SOP from Wiki
│
3. Agent invokes Python scripts → collects raw data
(Playwright → CDP Browser → parse → validate)
│
4. Agent analyzes data using DeepSeek-V4
→ applies Decision Framework
→ generates structured judgment layer
│
5. Agent constructs Feishu interactive card JSON
(schema: "2.0", elements[].tag: "markdown")
│
6. Agent calls send_feishu_card.py → pushes to group
│
7. Agent logs summary to memory/YYYY-MM-DD.md
→ optionally updates Wiki knowledge base
Schedule: Daily 10:30 CST
Source: Douyin Creator Platform (抖音创作者平台) - subscription page
Script: scripts/douyin_index.py
| Feature | Detail |
|---|---|
| Data Coverage | 8 venues: Jianye Movie Town + 7 core competitors |
| Metrics | Search Index (搜索指数) + Composite Index (综合指数) + Daily % Change |
| Collection | Dual-channel: Playwright script → CDP direct browser (auto-fallback) |
| Output | Ranked table with 8 venues, anomaly markers (🔺🔻) |
| Fallback | If script returns incomplete data (667-char truncation bug), agent auto-switches to CDP extraction |
| Real-World Case | May 22: script truncated → agent used CDP → report delivered on time |
Competitor Tracking Set (Fixed):
- 清明上河园 (Qingming Riverside Landscape Garden) - #1 by search volume
- 万岁山武侠城 (Wansui Mountain Martial Arts City) - #2
- 银基动物王国 (Yinji Animal Kingdom) - #3 (亲子旺季 surge)
- 郑州方特欢乐世界 (Zhengzhou Fantawild) - student pricing threat
- 郑州海昌海洋公园 (Zhengzhou Haichang Ocean Park) - price war
- 只有河南戏剧幻城 (Only Henan·Drama City) - 50%+ surge
- 只有红楼梦戏剧幻城 (Only Dream of Red Mansions) - growth leader
- 建业电影小镇 (Jianye Movie Town) - baseline
Schedule: Daily 10:00 CST Source: Xiaohongshu Lingxi Backend (小红书灵犀) + Search Page + Official Account Profile
| Feature | Detail |
|---|---|
| Official Account | 98K followers, 2,196 notes, profile ID: 5fbb4f740000000001000410 |
| Brand Metrics | Audience Assets, Search Volume, Click Index, Reading Penetration Rate |
| Competitor Keywords | 7 competitors + self brand - weekly comparison |
| Xiaohongshu Trend Data | Search totals, YoY/DoD change, Interest rankings |
| Viral Notes | Top UGC notes with like/collect/comment counts |
Schedule: Daily 13:00 CST Source: Baidu Search + Weibo Trending + Industry News Scope: National - NOT limited to 7 core competitors (expanded May 25)
Coverage Dimensions:
- 🏛️ Policy & Capital: Government tourism initiatives, subsidies, regulatory changes
- 🏟️ Competitor Activity: New shows, events, pricing changes, IP collaborations
- 🌐 Industry Trends: Summer tourism, experience economy, national travel patterns
⚠️ Risk Alerts: Safety incidents, regulatory enforcement, negative PR
Recent Examples (May 25):
- Henan Provincial Tourism Development Conference just concluded (5.22-23) with 300+ summer initiatives
- 5·19 China Tourism Day subsidy window closing May 31
- Ministry of Culture spot-slap on 5A downgrade threats (Shaolin Temple case)
Schedule: Daily 15:00 CST Source: Douyin Trending · Xiaohongshu Viral · Weibo Hot Search Scope: National - actively discovers new competitors
Deconstruction Framework:
Case: [Title]
├─ 📊 Data Snapshot (likes/views/engagement)
├─ 🔍 Strategy Analysis (angle/emotion/format)
├─ 🎯 Replicability Assessment (confidence score)
└─ 💡 Movie Town Adaptation (concrete action plan)
Recent Cases Analyzed:
- "Li Bai Poetry Duel at West Lake" - 356M views, NPC random-poetry model → replicable
- "Red Rose Dress for Waterfall" - Chongdugou scenic makeover → 22K likes
- "520 Marriage Registration" at Wansui Mountain - civic ceremony in scenic spot
- "Korean Brand Copies Hanfu" - Weibo #15 trending → cultural IP defense
Schedule: Daily 16:00 CST
Source: Douyin Index Keyword Page + Xiaohongshu Lingxi + Baidu Search
Method: Rotation mechanism with checkpoint resume (/tmp/daily_task_state.json)
4-Platform Collection Manifest:
| Platform | Data Collected | Tool |
|---|---|---|
| 🎵 Douyin Index | Search Index, Composite Index, Related Keywords TOP10, Audience Portrait | cdp_keyword_deep.py |
| 📕 Xiaohongshu Lingxi | Search Volume, Hot Terms, Upstream/Downstream Keywords | CDP + JS Injection |
| 📕 XHS Search Page | Viral Notes TOP10, Tags, Hot Questions | CDP Browser |
| 🌐 Baidu Search | Revenue, Visitor Count, Ticket Price, Media Coverage | CDP Browser |
Progress: 14/21 core competitors completed (as of May 25)
Schedule: Sundays 10:00 CST
Each week, the system calculates its own prediction accuracy across all decision outputs:
| Week | Accuracy | Trend |
|---|---|---|
| W18 | 60% | Baseline |
| W19 | 65% | +5pp |
| W20 | 70% | +5pp |
| W21 | 80% | +10pp |
Error Pattern Analysis (W21):
- Execution-layer disconnection: 2 errors (insights identified but execution not triggered)
- Data source breakage: 1 error (CSV missing 8 days)
- Signal misinterpretation: 0 errors (improving)
Schedule: Tuesday 9:30 CST (moved from Monday because Monday data is stale)
Data Source: ~/Desktop/2026游客量统计.csv - daily passenger count spreadsheet
Report Structure (5 chapters, max 5 tables per card):
- YTD Summary (年度累计 vs target)
- Monthly Breakdown (月度拆解)
- Last 7 Days Detail (近7日明细)
- 穿越德化街 Thematic Analysis (穿越德化街专项)
- Actionable Recommendations (建议)
Key Metrics (as of May 17):
| Month | Visitors | Days | Daily Avg | YTD Cumulative | Target Completion |
|---|---|---|---|---|---|
| Jan | 56,571 | 29 | 1,950 | 56,571 | 3.7% |
| Feb | 307,169 | 28 | 10,970 | 363,740 | 23.8% |
| Mar | 80,285 | 31 | 2,589 | 444,025 | 29.0% |
| Apr | 93,295 | 30 | 3,109 | 537,320 | 35.1% |
| May (to 17th) | 118,747 | 17 | 6,985 | 656,067 | 42.9% |
| Annual Target | 1,230,000 | - | - | - | - |
Data Quality Status:
| Source | Format | Location | Update Frequency |
|---|---|---|---|
| Primary: Daily spreadsheeet | 2026游客量统计.csv (wide-format, 368 cols) |
~/Desktop/ |
Irregular (internal dept) |
| Historical reference | 2023-2025年门票销售及客流统计数据表.xlsx |
~/Desktop/ |
Annual |
| Fallback: Feishu Bitable | 电影小镇-2026年数量统计 |
Feishu multi-dim table | Daily (when CSV stale) |
The CSV is a complex wide-format spreadsheet (not a clean row-per-day format). The parser (sync_obsidian_daily.py) handles:
Row Layout:
Row 0: 2023年参考 - 368 daily values (for YoY comparison)
Row 1: 2024年参考 - 368 daily values
Row 2: 2025年参考 - 32 values (partial year)
Row 3: 天气备注 - Weather notes per day
Rows 4-11: 门票细分 - Ticket breakdown by channel
Row 12: 门票人数合计 - Total tickets sold (primary metric)
Row 13: 门票收入金额 - Ticket revenue (¥)
Row 14: 闸机入园人次 - Turnstile entry count (actual visitors)
Rows 16-23: 预定数据 - Pre-sale data (evening/morning shifts)
Row 25-28: 穿越德化街 - Show performance data
Column mapping: Column index 2 = January 1, column 3 = January 2, etc. (sequential days from Jan 1).
CSV file detected → Python parser reads wide format
→ Extract: Total tickets (Row 12), Turnstile (Row 14), Weather (Row 3)
→ Build: {date: {tickets, turnstile, weather}} dictionary
→ Calculate: Daily avg, WoW change, YoY comparison, YTD cumulative
→ Compare: vs annual target (1,230,000), monthly benchmarks
→ Output: Structured data → Feishu card (weekly report)
→ Memory log (daily append)
→ Obsidian Wiki (data.md update)
~/Desktop/2026游客量统计.csv
│
├──→ scripts/sync_obsidian_daily.py (auto-detects changes)
│ │
│ ├──→ workspace wiki/电影小镇/历史数据/2026年/数据.md
│ │ (auto-updates last-updated timestamp)
│ │
│ └──→ Obsidian Vault (同步目标)
│ (mirrors workspace wiki to user's Obsidian)
│
└──→ Agent weekly passenger insight task (周二 9:30)
Reads CSV → generates 5-chapter card
→ pushes to Feishu group
→ logs to memory
| Metric | Calculation | Used For |
|---|---|---|
| Daily total | Row 12, col N | Base data point |
| Daily avg (month) | Monthly total ÷ days | Capacity planning |
| YTD completion | Cumulative ÷ 1,230,000 | Target tracking |
| WoW change | (This week - last week) ÷ last week | Trend detection |
| YoY change | (2026 - 2023/2024) ÷ baseline | Growth analysis |
| Channel breakdown | Rows 4-11 / total | Channel efficiency |
| Turnstile-to-ticket | Row 14 ÷ Row 12 | Redemption rate |
| Metric | Value |
|---|---|
| Last CSV update | 2026-05-17 |
| Data age | 8 days stale ( |
| YTD total | 656,067 |
| Target completion | 42.9% |
| Days with data | 135 days (Jan 1 - May 17) |
| Max single day | 33,411 (May 3, Labor Day) |
穿越德化街 is the flagship indoor theatrical performance at Jianye Movie Town - a time-travel immersive show set in 1930s Dehua Street, Zhengzhou. The venue underwent expansion in late 2024:
- Before expansion (2023): 450 seats per show
- After expansion (2025+): 1,140 seats per show (253% increase)
- Expansion impact: Oct-Nov 2024 closed for construction; Dec 2024 pressure-test only
The show data lives in the same passenger CSV (Rows 25-28):
| CSV Row | Data | Unit |
|---|---|---|
| Row 25 | Date label | String ("1月1日") |
| Row 26 | Performances (场次) | Count |
| Row 27 | Seat inventory (库存) | Seats (1,140 per show) |
| Row 28 | Tickets sold (售卖) | Tickets |
| (Derived) | Occupancy rate (上座率) | Sold ÷ Inventory |
1. Performance Cadence
- Regular days: 1 show/day (1,140 seats)
- Peak weekends: 2-3 shows/day
- Holiday peaks: 5-6 shows/day (Labor Day 2026: 23 shows in 5 days)
2. Occupancy Rate Tracking
Occupancy = Tickets Sold ÷ (Performances × 1,140)
Thresholds:
🔴 Below 40% → Under-performing, review pricing/content
🟡 40-65% → Normal range
🟢 65-85% → Good utilization
💎 Above 85% → Capacity constraint, consider adding shows
3. Conversion Rate (The Critical Metric)
Conversion Rate = Show Audience ÷ Park Visitors
Historical:
2023: 18.0% (baseline, pre-expansion)
2024: 16.2% (dipped, pre-expansion)
2025: 35.2% 🚀 (doubled post-expansion - product improvement)
2026 YTD: 27.0% (tracking - Q1 low due to off-season)
4. Ticket Mix (套票 vs 加购)
| Metric | 2023 | 2024 | 2025 | 2026 Q1 |
|---|---|---|---|---|
| Package ticket % (套票) | 72.5% | 60.7% | 77.6% | 49.5% |
| Add-on ticket % (加购) | 27.5% | 39.3% | 22.4% | 50.5% |
| Unit price (package) | - | - | ¥101.59 | - |
| Unit price (add-on) | - | - | ¥52.80 | - |
| Month | Days | Shows | Audience |
|---|---|---|---|
| April | 30 | 52 | 37,581 |
| May (to 17th) | 17 | 47 | 33,565 |
| Q2 Total | 47 | 99 | 71,146 |
May Labor Day Peak (May 1-4): 5-6 shows/day, 92-93% occupancy - near capacity.
CSV Rows 25-28 parsed → sync_obsidian_daily.py
→ Extracts: date, shows, inventory, sold
→ Calculates: occupancy %, conversion %, trend
→ Updates:
wiki/电影小镇/演出节目/穿越德化街.md
wiki/sources/穿越德化街数据分析.md
→ Weekly report: included in passenger insight card
| Source | Data | Update |
|---|---|---|
| Passenger CSV Row 13 | 门票收入金额 (Ticket revenue) | Per CSV sync |
| 2023-2025年门票销售及客流统计数据表.xlsx | Historical revenue | Annual |
| 穿越德化街 Excel | 演出收入 (Show revenue) | As available |
Total Revenue = Ticket Revenue + Non-Ticket Revenue
│ │
▼ ▼
× Daily Visitors Dining, Shopping,
× Average Ticket Accommodation,
Price (ATP) Photo, Experiences
Revenue Channels (Ticket):
| Channel | Description | Data Source |
|---|---|---|
| Online individual | 线上散客 (mini-program, Douyin, Meituan) | CSV Rows 9-10 |
| Offline window | 窗口散客 (walk-up at gate) | CSV Row 10 |
| Travel agency | 旅行社 (group tours) | CSV Row 8 |
| Corporate client | 大客户 (corporate events) | CSV Row 7 |
| Study tours | 研学 (educational groups) | CSV Row 6 |
| Year | Show Revenue | Note |
|---|---|---|
| 2023 | ¥3,139万 | Pre-expansion baseline |
| 2024 | ¥2,675万 | Pre-expansion, -14.8% |
| 2025 | ¥4,266万 | 🚀 Post-expansion, +35.9% |
| 2026 Q1 | ¥632万 | Tracking |
| Analysis | Method | Output |
|---|---|---|
| ATP (Avg Ticket Price) | Revenue ÷ Visitors | Pricing strategy input |
| Channel mix | Each channel ÷ total | Channel optimization |
| Non-ticket revenue share | (Total - Ticket) ÷ Total | Ancillary revenue tracking |
| Show revenue contribution | Show revenue ÷ Total | Product efficiency |
Every report, insight, and recommendation in this system follows a structured decision framework designed for real-world marketing execution.
Every daily report contains this exact judgment structure:
| Dimension | Values | Description | Example |
|---|---|---|---|
| 🎯 Impact Level | 🔴高/🟡中/🟢低/📡噪音 | How much does this affect Movie Town? | 🔴高 - competitor 50%+ surge |
| 💡 Action | 跟风/借势/警惕/忽略 | What should the team do? | 跟风 - replicate the format |
| ⏰ Window | 今天/本周/不紧急 | When must action happen? | 今天 - trending window closes in 48h |
| Specific loss statement | What is lost by doing nothing? | Missed summer peak season |
D<number> | <one-line title>
├─ 简单理解: plain-English explanation
├─ 做错风险: consequence of not acting
├─ 推荐+理由: specific suggestion + confidence score
└─ 利弊: trade-offs (benefit vs cost/effort)
The agent is strictly prohibited from using these vague phrases:
| ❌ Banned | ✅ Replace With |
|---|---|
| "值得关注" | "本周必须执行" or specific priority |
| "仅供参考" | Concrete recommendation with confidence |
| "或许可以考虑" | Definitive suggestion: "推荐:..." |
| "需要进一步分析" | Clear judgment or explicit "信息不足" |
Every Sunday, the system:
- Calculates prediction accuracy from the past week
- Analyzes error patterns: data gaps / logic bias / platform changes
- Updates decision rules in Wiki
- Validates → marks rules as 🟢 (verified) or 🔴 (expired)
The wiki follows a three-layer knowledge abstraction model:
KNOWLEDGE LAYER (wiki/)
│
├── concepts/ ← Abstract patterns & theories
│ ├── 演艺景区.md - Performance venue characteristics
│ ├── 内容爆款规律.md - Viral content patterns
│ ├── 景区营销漏斗.md - Marketing funnel model
│ ├── 情绪营销.md - Emotional marketing framework
│ ├── 平台算法规则.md - Douyin/Xiaohongshu algorithm notes
│ └── ... (12 concept files)
│
├── entities/ ← Concrete objects & actors
│ ├── 建业电影小镇.md - Movie Town profile
│ ├── 万岁山武侠城.md - Competitor: Wansui Mountain
│ ├── 清明上河园.md - Competitor: Qingming Garden
│ ├── 抖音平台.md - Platform entity
│ ├── 小红书平台.md - Platform entity
│ └── ... (12 entity files)
│
├── sources/ ← Data provenance & analysis
│ ├── 穿越德化街数据分析.md - Thematic data analysis
│ ├── 抖音指数追踪日报.md - Daily tracking archive
│ ├── 竞品深度档案.md - Profile sources
│ └── ... (8 source files)
│
└── queries/ ← Archived decision Q&A
├── 抖音与小红书平台差异.md
├── 知识层与业务层关系.md
└── ... (5 query files)
BUSINESS LAYER (wiki/电影小镇/)
│
├── 基础档案.md - Basic info, targets, annual goals
├── 战略框架.md - SWOT, competitive positioning
├── 人群画像.md - Douyin & Xiaohongshu audience profiles
├── 历史数据/ - Historical records (2023-2026)
│ ├── 2023年/数据.md
│ ├── 2024年/数据.md
│ ├── 2025年/数据.md
│ └── 数据.md (2026, updated to May 17)
├── 演出节目/
│ └── 穿越德化街.md - 6-year show data + Q2 update
├── 运营方法/ - Operation SOPs
│ ├── 抖音运营方法.md
│ └── 小红书运营方法.md
└── 运营规划/ - Seasonal plans
└── 环形水剧场与复古广场夏季运营方案.md
COMPETITOR INTELLIGENCE (wiki/竞品分析/)
│
├── 竞品深度档案/ - 20 deep-profile reports
│ ├── 郑州方特欢乐世界深度分析.md
│ ├── 银基动物王国深度分析.md
│ ├── 清明上河园深度分析.md
│ ├── 大唐不夜城深度分析.md
│ ├── 阿那亚深度分析.md
│ └── ... (20 files, all completed)
│
├── 竞品动态追踪/ - Daily activity log (Apr 20-27)
├── 追踪数据/
│ ├── 抖音指数追踪.md
│ └── 小红书爆款追踪.md
│
└── 关键词池状态.md - Checkpoint system for rotation
CASE LIBRARY (wiki/全国景区案例库/)
│
├── index.md - 10 viral formulas index
└── (20+ weekly case files)
├── 大唐不夜城夜游标杆-2026W17.md
├── 万岁山武侠城标杆-2026W17.md
├── 打铁花跨景区爆款现象-2026W17.md
├── 乌镇住宿早茶客46%复购率-2026W22.md
├── NPC体系化运营青岩古镇银票系统-2026W22.md
└── ...
10 Viral Formulas (as of W22):
| # | Formula | Example Cases |
|---|---|---|
| 1 | Emotional Teaser + Reality Reveal | Wansui Mountain, Tang Dynasty Street |
| 2 | Intangible Heritage × Scenic Reality | Iron Flower, Song Dynasty Performance |
| 3 | Local KOC Matrix Distribution | Multi-angle coverage strategy |
| 4 | Interactive NPC Surprise Encounters | Poetry Duel, Marriage Registration |
| 5 | Seasonal Limited-Edition Transformation | Summer Night Market, Cherry Blossom |
| 6 | Reverse Operation: Quiet ≠ Boring | Only Henan "Non-Noisy" Concert |
| 7 | Spirituality Economy × Emotional Bonding | First-snow Wishes, Sunset Photography |
| 8 | Space Cultural Remake | Waterfall Rose Dress, Cliff Coffee |
| 9 | Festival IP Long-term Operation | Qingming Festival series, Dragon Boat |
| 10 | Audience Co-creation UGC Campaign | "My Hidden Spot" Contest |
| # | SOP File | Purpose |
|---|---|---|
| 1 | 抖音指数日报.md | Daily Douyin report generation flow |
| 2 | 小红书日报.md | Daily XHS report generation flow |
| 3 | 文旅活动热点追踪日报.md | Travel intel daily report |
| 4 | 竞品爆款拆解.md | Viral hit deconstruction standard |
| 5 | 竞品关键词深度分析流程.md | Keyword deep analysis step-by-step |
| 6 | 竞品内容动态.md | Competitor content tracking |
| 7 | 竞品深度分析流程.md | 4-platform deep analysis SOP |
| 8 | 竞品深度档案标准格式.md | Profile archive format spec |
| 9 | 周度客流营收洞察报告.md | Weekly passenger insight report |
| 10 | 每日复盘整合.md | Daily review integration |
| 11 | 每日任务总览.md | All tasks roster |
| 12 | 案例库更新.md | Case library update flow |
| 13 | 系统健康检查SOP.md | System health checklist |
| 14 | 专属浏览器维护SOP.md | CDP browser maintenance |
| 15 | 飞书卡片视觉规范.md | Feishu card visual spec |
| 16 | 飞书卡片故障复盘.md | Card failure postmortem |
| 17 | 反馈纠错.md | Feedback & correction |
| 18 | 日报模板.md | Report template |
| 19 | Wiki健康检查.md | Wiki health check |
| 20 | 代码库Wiki漂移检查.md | Wiki drift detection |
| 21 | 决策简报格式标准.md | Decision brief format |
| 22 | 反谄媚分析规范.md | Anti-sycophancy analysis norm |
memory/ ← Agent's episodic memory
├── YYYY-MM-DD.md ← Daily logs (cron execution records)
├── topics/feedback/ ← Correction/confirmation feedback
├── topics/projects/ ← Long-running project state
├── heartbeat-state.json ← Heartbeat check tracking
└── daily_task_state.json ← Task rotation checkpoint
MEMORY.md ← Long-term curated memory (100 lines max)
PROGRAMMER_AGENT.md ← Dedicated coding agent identity
AGENTS.md ← Agent behavioral guidelines
SOUL.md ← Agent personality & working standards
TOOLS.md ← Local tool configuration
USER.md ← User profile & preferences
| Rule | Purpose | Established |
|---|---|---|
Feishu cards must use send_feishu_card.py |
Tables won't render via default message tool | 2026-05-25 |
| Weekly passenger report moved to Tuesday | Monday data is stale | 2025-05-25 |
| Obsidian sync: value-add only | Don't re-sync unchanged content | 2025-05-25 |
| Search scope: unlimited national | Any relevant competitor qualifies | 2025-05-25 |
| Data must be read from actual files | Never use experience/heuristics | 2025-04-22 |
| browser-use is banned entirely | All automation via Playwright scripts | 2025-04-20 |
| Cron delivery: mode=none | No redundant push notifications | 2025-04-10 |
Card line breaks: use <br/> |
Feishu doesn't parse \n |
2025-04-10 |
The system uses /tmp/daily_task_state.json for task rotation persistence:
{
"竞品关键词": {
"done": ["万岁山武侠城", "清明上河园", "只有河南", "郑州方特", "银基", ...],
"current": "郑州海昌海洋公园"
},
"文旅案例": {
"done": ["万岁山", "银基", "大唐不夜城", ...],
"current": "轮换中"
}
}Port: 18800 (dedicated browser instance)
Tab Assignments:
Tab 0: Xiaohongshu Lingxi Backend (idea.xiaohongshu.com/trend/trendAnalyze)
Tab 1: Baidu Search
Tab 2: Douyin Subscription Page (creator.douyin.com/my-subscript)
Tab 3: Douyin iframe
Tab 4: Douyin Keyword Page (creator.douyin.com/arithmetic-index)
Tab 5: Douyin iframe
Tab 6: Xiaohongshu Explore Page (xiaohongshu.com/explore)
Cookies stored at:
/tmp/juLiang_cookies.json - Douyin (proxied via 127.0.0.1:7897)
/tmp/xiaohongshu_cookies.json - Xiaohongshu (proxied)
/tmp/weibo_cookies.json - Weibo
| Option | Considered | Verdict |
|---|---|---|
| OpenClaw + Agent Skills | ✅ Chosen | 25+ model providers, cron scheduling, skills ecosystem, local-first |
| Custom Python scripts only | ❌ Rejected | No agent reasoning or autonomous decision-making |
| Paid SaaS (e.g. Similarweb) | ❌ Rejected | Expensive, no custom decision framework, data locked-in |
WeChat Mini-Programs cannot run npm d3-force directly (ES module + DOM dependency). The graph visualization uses a hand-rolled Verlet integration physics engine:
const kRepulsion = 700; // Coulomb-like repulsion
const kSpring = 0.04; // Hooke's spring tension
const lLength = 120; // Ideal spring length
const gravity = 0.02; // Center gravity
const friction = 0.85; // Velocity dampingPerformance: 60fps on modern devices with 50+ nodes.
The Douyin Creator Platform uses heavy client-side rendering (SPA). The Playwright script sometimes returns empty or truncated data. The CDP direct browser channel serves as a reliable fallback:
Script → [success?] → Parse & use data
↓
[fail?] → CDP direct extraction → Parse & use data
↓
[fail?] → Trend inference from recent history
Feishu's default message tool sends msg_type: "post" (rich text), which does not render tables. The card script sends msg_type: "interactive" with proper schema: "2.0" format:
# send_feishu_card.py
payload = {
"receive_id": chat_id,
"msg_type": "interactive",
"content": json.dumps(card, ensure_ascii=False)
}
# card format:
{
"schema": "2.0",
"header": {"title": {"tag": "plain_text", "content": "..."}},
"body": {"elements": [{"tag": "markdown", "content": "| table | data |"}]}
}Beyond the core modules, the system includes a rich ecosystem of supporting tools, operational mechanisms, and quality assurance processes.
The system includes a comprehensive Python scripts library that handles everything from data collection to card delivery.
| Script | Purpose | Type | Key Feature |
|---|---|---|---|
douyin_index.py |
Douyin Index data collection | 📊 Data | Dual-channel (Playwright + CDP) |
competitor_keyword_v8.py |
Competitor keyword deep analysis | 📊 Data | 4-platform aggregation |
cdp_cookie_hub.py |
Cookie extraction from CDP browser | 🔧 Ops | Cross-platform (Douyin/XHS/Weibo) |
cdp_keyword_deep.py |
Deep keyword analysis via CDP | 📊 Data | Auto-rotate between Tab 2/4 |
cdp_collect.py |
General CDP data collector | 📊 Data | Tab-aware targeting |
xiaohongshu_crawl.py |
Xiaohongshu data collection | 📊 Data | Anti-ratelimit handling |
xhs_competitor_crawl.py |
XHS competitor page scraping | 📊 Data | Multi-keyword queue |
send_feishu_card.py |
Feishu interactive card sender | 📨 Push | schema 2.0 validation |
sync_obsidian_daily.py |
Obsidian Wiki sync agent | 🔄 Sync | CSV→Wiki→Vault pipeline |
query_passenger.py |
Passenger CSV query helper | 📊 Data | Date-range filtering |
self_check.py |
System self-diagnosis | 🔧 Ops | 8-point health checklist |
cdp_restore_tabs.py |
CDP browser tab recovery | 🔧 Ops | Auto-restore on crash |
case_library_scan.py |
Case library audit & repair | 🔄 Sync | Broken link detection |
industry_news_browser.py |
Travel intel news collector | 📊 Data | Multi-source aggregation |
llmwiki_ingest.py |
Wiki knowledge ingestion | 🔄 Sync | karpathy-wiki pattern |
build_xhs_card.py |
XHS report card builder | 🎨 Card | Visual spec compliant |
build_dashboard.py |
|||
validate_data.py |
Data integrity validator | ✅ QA | Staleness & completeness |
project_drift_check.py |
Project drift detection | ✅ QA | Content vs. reality mismatch |
wiki_drift_check.py |
Wiki drift detection | ✅ QA | Cross-reference validation |
Total: 45+ active scripts (archived/legacy excluded)
All reports are delivered as Feishu interactive cards (not plain text). The card system has its own design specification, validation script, and fault recovery.
{
"schema": "2.0", // Required for rendering
"header": {
"title": {
"tag": "plain_text", // NOT "lark_md"
"content": "📊 抖音指数日报 | 2026-05-25"
},
"template": "blue" // Optional color accent
},
"body": {
"elements": [ // NOT root-level elements[]
{
"tag": "markdown",
"content": "| 景区 | 搜索指数 | ... |" // Tables inside markdown
}
]
}
}
| Rule | Reason | Established |
|---|---|---|
header.title.tag must be "plain_text" |
"lark_md" causes rendering glitch |
2026-04-19 |
Content goes in body.elements[], not root elements[] |
Schema compliance | 2026-04-19 |
Tables inside markdown elements with | pipe syntax |
Feishu MD renderer | 2026-04-19 |
Line breaks use <br/> not \n |
Feishu ignores \n |
2026-04-10 |
Must use send_feishu_card.py, not default message tool |
msg_type:"post" doesn't render tables |
2026-05-25 |
Section separator: ━━━━━━━━━━━━━━━━━━━━━━━━
Emoji hierarchy: 📌 → 🔍 → ⚠️ → 💡
Bold for emphasis: **text**
Code blocks: `monospace text`
Decision output: D<number> | <one-line title>
# send_feishu_card.py (key logic)
# 1. Validate card structure
validate_card(card) → checks schema/header/elements
# 2. Get Feishu tenant token
token = get_token() # client_credentials grant
# 3. Send with correct msg_type
payload = {
"receive_id": chat_id,
"msg_type": "interactive", # NOT "post"
"content": json.dumps(card, ensure_ascii=False)
}
# 4. Auto-retry on token expiry
if result.code in (99991663, 19001):
token = refresh_token()
retry()| Date | Failure | Root Cause | Fix |
|---|---|---|---|
| 05-25 | W22 marketing calendar shown as plain text | Agent used default message tool (msg_type: "post") instead of send_feishu_card.py |
Updated all 9 cron jobs to enforce card script |
| 04-27 | Card table not rendering | header.title.tag was "lark_md" |
Changed to "plain_text" |
| 04-19 | Elements in root instead of body.elements[] |
Agent misread schema spec | Added validation in send_feishu_card.py |
The system uses a multi-layer fallback strategy to ensure uninterrupted daily operations:
Layer 1: Primary Script
└─ douyin_index.py runs → [success?]
├─ Yes → parse + use data
└─ No → fall to Layer 2
Layer 2: CDP Direct Extraction
└─ CDP browser navigates to target page → [success?]
├─ Yes → parse + use data
└─ No → fall to Layer 3
Layer 3: Trend Inference
└─ Estimate from recent history + known patterns
→ mark as "estimated" (lower confidence)
| Error Pattern | Trigger | Response | Recovery |
|---|---|---|---|
| 503 Service Busy | DeepSeek API overload | Auto-retry at next cron cycle | No action needed (provider) |
| Cookie Expired | Douyin/XHS session timeout | CDP browser fallback; next sync cycle | Re-login via QR code |
| 667-char Truncation | Douyin SPA page partial render | Auto-switch to CDP extraction | Script update pending |
| Tool Execution Timeout | CDP browser busy/heavy load | Agent retries with 30s timeout | Monitor, no manual fix |
| Edit Failed | Wiki file write conflict | Agent retries with new content | Rare, auto-resolves |
Alert configuration per cron job (example):
{
"failureAlert": {
"after": 2, // Alert after 2 consecutive failures
"channel": "feishu",
"to": "oc_f109bcfd1bc7e166fd0ae077f70247cf",
"cooldownMs": 60000 // 1 hour between alerts
}
}The agent's capabilities are extensible through OpenClaw's Skills system. Skills are loaded from ClawHub (community registry) or custom-written.
| Skill | Purpose | Installed | Type |
|---|---|---|---|
wechat-mini-program-builder |
WeChat mini-program rapid dev | ✅ | 📱 Dev |
mini-program-dev |
Mini-program code templates & API | ✅ | 📱 Dev |
wechat-miniprogram-skill |
Mini-program beginner→expert guide | ✅ | 📱 Dev |
miniprogram-development |
General mini-program dev | ✅ | 📱 Dev |
frontend-design-3 |
Frontend design specification | ✅ | 🎨 UI |
react-best-practices |
React best practices | ✅ | 💻 Code |
typescript-skills |
TypeScript skill set | ✅ | 💻 Code |
karpathy-coding-guidelines |
Karpathy coding principles | ✅ | 💻 Code |
debug-pro |
Systematic debugging | ✅ | 🔧 Dev |
api-tester |
HTTP request testing (GET/POST/PUT/DELETE) | ✅ | 🔧 Dev |
browser-automation |
Browser automation via natural language | ✅ | 🤖 Auto |
karpathy-guidelines |
General LLM coding wisdom | ✅ | 💻 Code |
Total skills available: 59 (system + workspace + installed)
Skill = SKILL.md (YAML frontmatter + Markdown instructions)
→ Agent reads skill at load time
→ Matches skill description against user request
→ Activates relevant skills
→ Filters by environment/config
Priority: Workspace > Local > Bundled
- Registry: 13,729 community-built skills (as of Feb 2026)
- Installer:
clawhub install <skill-slug> --dir ~/.openclaw/workspace/skills - Security: VirusTotal integration for published skills
The agent maintains both episodic memory (daily logs) and semantic memory (curated rules), plus a feedback loop for continuous improvement.
EPISODIC MEMORY (Raw logs)
memory/
├── YYYY-MM-DD.md ← Daily execution records
├── YYYY-MM-DD.md.bak ← Archived (when compacted)
SEMANTIC MEMORY (Curated)
MEMORY.md ← Long-term rules (max 100 lines)
├── 铁律 (Ironclad rules) — Violation must-correct
├── 关键洞察 (Key insights) — Reusable patterns
├── [reference] — System pointers (Feishu groups, file paths)
├── [project] — Active project status
└── [feedback] — Confirmed corrections/confirmations
STATE (Machine-readable)
memory/heartbeat-state.json ← Heartbeat check tracking
memory/topics/ ← Topic-specific knowledge
├── feedback/ ← User corrections & confirmations
├── projects/ ← Long-running project state
└── daily-tasks.md ← Task roster
User says "不要" / "不对" / "停止"
→ Agent records correction in memory/topics/feedback/
→ Updates MEMORY.md rules if pattern confirmed
→ Adjusts future behavior
User says "对" / "很好" / "就这样"
→ Agent records confirmation (success pattern)
→ Reinforces existing rule
→ No change needed
**规则：** [Brief rule description]
**Why:** [Why this rule exists]
**How to apply:** [When/where to apply]| Type | Purpose | Example |
|---|---|---|
[user] |
User role/preferences/goals | user: 站长偏好详细数据报告 |
[feedback] |
Work guidance (correction+confirmation) | feedback: 达人必须绑定转化链路 |
[project] |
Project status/targets | project: 当前在优化多Agent系统 |
[reference] |
External system pointers | reference: 飞书群 oc_xxx |
| Rule | Type | Established |
|---|---|---|
| Weekly passenger report moved to Tuesday | 铁律 | 2026-05-25 |
Feishu cards must use send_feishu_card.py |
铁律 | 2026-05-25 |
| Search scope: unlimited national (not 21 fixed) | 铁律 | 2026-05-25 |
| Data must be read from actual files, never guessing | 铁律 | 2026-04-22 |
| browser-use is banned (use Playwright scripts) | 铁律 | 2026-04-20 |
| Cron delivery mode: none (no redundant announce) | 铁律 | 2026-04-10 |
The system self-diagnoses every 30 minutes via heartbeat:
Checklist:
□ Cron jobs ran successfully (check lastError)
□ CDP browser online (port 18800)
□ Cookie files fresh (not expired)
□ Passenger CSV not stale (last update < 14 days)
□ Feishu Bot token valid
□ Disk space adequate (< 90% usage)
□ No orphan session files (> 100 = cleanup needed)
□ Skills & plugins loaded without errors
| Task | Time | Description |
|---|---|---|
| Wiki Health Check | 10:00 | Verify all wiki links, fix broken refs |
| Codebase Drift Check | 10:00 | Detect workspace vs wiki content drift |
| Orphan Session Cleanup | 11:00 | Archive/deleted unused .jsonl transcript files |
| Skill Exploration | 14:00 | Discover new ClawHub skills, update skillset |
| Weekly Evolution Review | Weekly | System upgrade, memory consolidation, SOP audit |
| Property | Detail |
|---|---|
| Port | 18800 (dedicated instance) |
| Target | host (Mac Mini) |
| Tab 0 | Xiaohongshu Lingxi Backend (trend/trendAnalyze) |
| Tab 1 | Baidu Search |
| Tab 2 | Douyin Subscription Page (my-subscript) |
| Tab 3 | Douyin iframe |
| Tab 4 | Douyin Keyword Page (arithmetic-index) |
| Tab 5 | Douyin iframe |
| Tab 6 | Xiaohongshu Explore Page (explore) |
| Cookie Storage | /tmp/juLiang_cookies.json (Douyin) |
/tmp/xiaohongshu_cookies.json (XHS) |
|
/tmp/weibo_cookies.json (Weibo) |
|
| Proxy | 127.0.0.1:7897 (for Douyin) |
| Recovery | cdp_restore_tabs.py on tab crash |
Cron: 每日 08:05 (cdp_cookie_hub.py)
→ Connect to CDP browser (port 18800)
→ Navigate each platform tab
→ Extract cookies via document.cookie
→ Save to /tmp/<platform>_cookies.json
→ (Fallback: stale cookies still work for ~24h)
| Gate | Check | Action on Failure |
|---|---|---|
| ✅ | Script returns all 8 venues | Partial data → switch to CDP |
| ✅ | Cookie age < 24h | Stale → CDP fallback, flag for refresh |
| ✅ | CSV last update < 2 weeks | Stale → flag in weekly report, request sync |
| ✅ | Feishu card delivery confirmed | Fail → retry with token refresh |
| ✅ | Wiki write succeeds | Conflict → retry with unique content |
Every Sunday, the system undergoes a structured evolution cycle:
| Phase | Action | Output |
|---|---|---|
| 🧹 Cleanup | Archive orphan sessions, compact memory | Clean state |
| 📚 Ingest | Process new wiki content (raw/ → knowledge layer) | Updated wiki |
| 🔍 Audit | Check prediction accuracy, error patterns | Accuracy report |
| 🧠 Learn | Update decision rules based on validation results | Rule updates |
| 🚀 Explore | Search ClawHub for new useful skills | Skill updates |
| 📝 Commit | Git commit & push Wiki updates | GitHub sync |
| Resource | Usage | Management Strategy |
|---|---|---|
| API Tokens | ~200K tokens/day (DeepSeek-V4-Flash) | Single model to max context window; isolated sessions prevent state bloat |
| Disk | 81Gi available / 228Gi total | Weekly orphan cleanup; Ollama models removed (reclaimed 26Gi) |
| CDP Browser | 6 permanent tabs | Tab-specific targeting prevents resource waste |
| Cron Sessions | 23 isolated sessions | Ephemeral (deleted after run); 30s-600s timeout per task |
| Feishu API | ~30 calls/day | Token caching reduces auth requests |
| Metric | Value |
|---|---|
| Uptime | 30+ days continuous |
| Daily reports generated | 6-8 decision briefs |
| Daily API tokens consumed | ~200K (DeepSeek-V4-Flash) |
| Feishu cards pushed | 180+ cards to date |
| Automated cron tasks | 23 active |
| Models allowed | deepseek/deepseek-v4-flash (single) |
| Asset Type | Count | Details |
|---|---|---|
| Competitor deep profiles | 20 | Full 4-platform analysis |
| Viral case studies | 20+ | Weekly updated |
| Viral formulas | 10 | Identified patterns |
| SOP documents | 22 | Standardized procedures |
| Historical records | 4 years | 2023-2026 passenger data |
| Wiki files | 150+ | Markdown documents |
- Search Index recovery: Movie Town search index rose +24.51% in W21 (brand trauma recovery signal)
- Competitive ranking: Moved from 7th to 4th among 8 tracked venues
- Weekly accuracy: Improved from 60% (W18) to 80% (W21) - 4 consecutive weeks of improvement
- Content vacuum detection: Algorithm successfully identified "search up / composite down" divergence pattern (content supply gap), confirmed May 22
| Layer | Technology | Purpose |
|---|---|---|
| AI Framework | OpenClaw v2026.5.20 | Agent runtime, cron, tools |
| LLM | DeepSeek-V4-Flash | Reasoning, analysis, generation |
| Data Collection | Python 3.12 + Playwright + CDP | Web scraping, browser automation |
| Push Channel | Feishu Bot API (interactive card) | Report delivery to group chat |
| Knowledge Base | Obsidian + karpathy-wiki | Structured wiki with 3-layer abstraction |
| Operating System | macOS 26.4 (Darwin 25.4.0) ARM64 | Host environment |
| Hardware | Mac Mini (2024) | 24/7 local runtime |
| Node.js | v25.8.2 | OpenClaw runtime |
| Skills Registry | ClawHub (13,729 available skills) | Extensible agent capabilities |
| Version Control | Git + GitHub | Code & wiki repository |
Open the wiki/ directory in Obsidian. Start at wiki/index.md.
All reports auto-push to Feishu group chat. Summary logs are in memory/YYYY-MM-DD.md.
scripts/- Python automation scripts (45+ files)wiki/- Knowledge base (150+ markdown files)PROGRAMMER_AGENT.md- Dedicated coding agent
The system includes a feedback loop mechanism. Corrections are recorded in memory/topics/feedback/ and influence future behavior.
Built by AI, for humans. Running since April 2026. Automatically deployed & maintained by OpenClaw AI Agent. Last system update: 2026-05-25
The system is under active development. The following capabilities are planned or in progress:
| Feature | Status | Description |
|---|---|---|
| 🔴 Viral Video Auto-Deconstruction | 🟡 Planning | Automated scraping of Douyin trending videos → AI analysis → replicability assessment → Feishu card. Currently manual-selection. Target: fully automated daily pipeline. |
| 🔴 Passenger Flow Real-Time Dashboard | 🟡 Planning | Replace stale-CSV dependency with real-time API from park ticketing system. Auto-detect anomalies (daily deviation > 20%). |
| 🟣 Revenue Tracking Module | 🔴 Research | Integrate revenue data from ticketing system + show system. Automated ATP calculation, revenue mix analysis, non-ticket revenue tracking. |
| 🟣 Competitor Pricing Monitor | 🔴 Research | Daily auto-check of competitor ticket prices (Douyin团购, Meituan, official mini-program). Alert on price changes > 10%. |
| Feature | Status | Description |
|---|---|---|
| 🟢 Weekly Prediction Accuracy | ✅ Active | Already implemented (W21: 80%). Continuous improvement via learning loop. |
| 🟣 Sentiment Analysis on UGC | 🔴 Research | Auto-analyze sentiment trends on Xiaohongshu/Douyin comments. Early warning for negative sentiment spikes. |
| 🟣 Automated Content Generation | 🔴 Research | Generate draft content (short videos scripts, Xiaohongshu posts) based on trending formats. Human review before publishing. |
| 🟣 Multi-Agent Collaboration | 🟡 Planning | Specialized sub-agents: Competitor Agent, Content Agent, Passenger Agent, Review Agent - working in parallel under coordinator. |
| Feature | Status | Description |
|---|---|---|
| 🟣 Public API Layer | 🔴 Research | Expose anonymized competitive intelligence data via API for other scenic spots. |
| 🟣 Template Marketplace | 🔴 Idea | Shareable SOP templates, decision frameworks, and report card templates. |
| 🟣 Cross-Scene Benchmarking | 🔴 Idea | Compare Movie Town metrics against national averages (seasonally adjusted). |
| 🟣 WeChat Mini-Program Extension | 🟡 In Dev | ChatWiki mini-program (holographic knowledge graph + AI ingestion center) - currently in active development. |
| Icon | Meaning |
|---|---|
| 🟢 | Active - deployed and running |
| 🟡 | Planning - design/spec in progress |
| 🔴 | Research - feasibility study |
| 💡 | Idea - concept, no active work |
Scenic-Area-Marketing-CN 是一个基于 OpenClaw AI Agent框架 构建的生产级全自动景区营销情报系统。系统 24/7 运行在 Mac Mini 上,自主完成从多平台数据采集、结构化竞争情报分析、决策级报告生成、飞书群推送,到 Obsidian 知识图谱归档的完整闭环。
系统服务于建业电影小镇--位于河南郑州的文化旅游景区,2026年度目标客流123万人次（2026-08-13下调，原153万）、营收1.2亿元。
这不是一个演示版或原型。系统已连续无人工干预运行30+天,日均产出 6-8份可执行决策简报。
┌─────────────────────────────────────────────────────────────────────────────────────┐
│ 🧠 AI AGENT 核心 │
│ OpenClaw Runtime · DeepSeek-V4-Flash · 23个cron任务 │
│ Web搜索 · 记忆系统 · CDP浏览器 · 飞书API │
└─────────────────────────────────────────────────────────────────────────────────────┘
│ │ │
▼ ▼ ▼
┌─────────────────────┐ ┌─────────────────────────────┐ ┌─────────────────────────┐
│ 📊 数据层 │ │ 🧠 智能分析层 │ │ ⚡ 执行层 │
├─────────────────────┤ ├─────────────────────────────┤ ├─────────────────────────┤
│ 抖音指数追踪 │ │ 决策判断引擎 │ │ 飞书卡片推送 │
│ 小红书监测 │ │ 竞品情报分析 │ │ (send_feishu_card.py) │
│ 微博热搜 │ │ 趋势&异动检测 │ │ Cron任务调度器 │
│ 百度搜索 │ │ 爆款拆解框架 │ │ Cookie轮换管理器 │
│ 客流CSV读取 │ │ 周度预测&复盘 │ │ CDP浏览器控制 │
│ CDP Cookie管理 │ │ 学习闭环&验证 │ │ 健康监控 │
└─────────────────────┘ └─────────────────────────────┘ └─────────────────────────┘
│ │ │
└─────────────────────────────┼───────────────────────────┘
▼
┌─────────────────────────────┐
│ 📚 知识层 │
├─────────────────────────────┤
│ Obsidian Wiki (三层抽象) │
│ → 概念/实体/来源 │
│ → 20个竞品深度档案 │
│ → 10条爆款公式/22个SOP │
│ → 历史数据 (2023-2026) │
└─────────────────────────────┘
┌──────────┐ ┌──────────────┐ ┌───────────────┐ ┌────────────┐
│ 08:00 │───▶│ 10:00-10:30 │───▶│ 13:00-16:00 │───▶│ 18:00-22:00│
│ Cookie │ │ 抖音+小红书 │ │ 情报+爆款 │ │ 复盘+ │
│ 同步 │ │ 日报 │ │ 深度分析 │ │ 归档 │
└──────────┘ └──────────────┘ └───────────────┘ └────────────┘
│ │ │ │
▼ ▼ ▼ ▼
┌──────────┐ ┌──────────────┐ ┌───────────────┐ ┌────────────┐
│ 3个Cookie │ │ 解析数据+ │ │ 决策简报 │ │ 记忆+Wiki │
│ 文件 │ │ 决策层输出 │ │ (D1-D4) │ │ +GitHub │
│ (抖音/小红 │ │ │ │ +风险预警 │ │ 提交 │
│ 书/微博) │ │ │ │ │ │ │
└──────────┘ └──────────────┘ └───────────────┘ └────────────┘
│ │
▼ ▼
┌──────────────────────────────┐
│ 飞书群聊 │
│ (oc_2581c03b79e4893cc36...) │
│ 通过send_feishu_card.py │
│ 发送interactive卡片 │
└──────────────────────────────┘
1. Cron触发隔离Agent session
│
2. Agent读取任务描述及Wiki中的SOP
│
3. Agent调用Python脚本 → 采集原始数据
(Playwright → CDP浏览器 → 解析 → 验证)
│
4. Agent用DeepSeek-V4分析数据
→ 应用决策判断框架
→ 生成结构化决策层
│
5. Agent构造飞书交互卡片JSON
(schema: "2.0", elements[].tag: "markdown")
│
6. Agent调用send_feishu_card.py → 推送至群
│
7. Agent记录执行摘要至memory/YYYY-MM-DD.md
→ 可选更新Wiki知识库
时间: 每日10:30
来源: 抖音创作者平台·我的订阅页
脚本: scripts/douyin_index.py
| 特性 | 详情 |
|---|---|
| 数据范围 | 8个景区:电影小镇+7核心竞品 |
| 指标 | 搜索指数 + 综合指数 + 日环比 |
| 采集 | 双通道:Playwright脚本 → CDP浏览器直连(自动降级) |
| 输出 | 排名表 + 异动标注(🔺🔻) |
| 降级机制 | 脚本返回空/截断数据时,自动切换CDP提取 |
8个常规定义竞品(固定不变):
| 排名 | 景区 | 日均搜索指数 | 特点 |
|---|---|---|---|
| 🥇 | 清明上河园 | 31.5万 | IP联动(杨洋/雨霖铃) |
| 🥈 | 万岁山武侠城 | 5.5万 | 情绪价值内容/IP势能 |
| 🥉 | 银基动物王国 | 9.4万 | 亲子旺季启动 |
| 4 | 郑州方特欢乐世界 | 1.1万 | 学生低价促销 |
| 5 | 郑州海昌海洋公园 | 5,557 | 价格战 |
| 6 | 只有河南戏剧幻城 | 5,162 | 50%+暴涨 |
| 7 | 只有红楼梦戏剧幻城 | 3,509 | 增速领先 |
| - | 建业电影小镇 | 8,448 | 基准 |
时间: 每日10:00 来源: 小红书灵犀后台 + 搜索页 + 官方账号
| 特性 | 数据 |
|---|---|
| 官方账号 | 9.8万粉丝 / 2196篇笔记 |
| 品牌五维 | 人群资产+44.64%, 搜索量+22.9%, 阅读渗透率↓57.93% |
| 决策输出 | 520方向正确 → 跟风提炼至618/端午 |
| 竞品覆盖 | 清上河园/万岁山/银基/只有河南/方特 |
时间: 每日13:00 范围: 全国范围(5月25日起扩展,不限7个核心竞品)
覆盖维度:
- 🏛️ 政策资本:政府旅游政策/补贴/监管变化
- 🏟️ 竞品活动:新节目/活动/定价/IP合作
- 🌐 行业趋势:暑期旅游/体验经济/出行模式
⚠️ 风险预警:安全事故/监管执法/负面舆情
今日(5月25日)示例:
- 河南省文旅发展大会刚闭幕(5.22-23安阳),发布300项惠民措施
- 5·19中国旅游日惠民补贴窗口5月31日关闭
- 文旅部第二批强制消费典型案例涉及河南新乡/洛阳
时间: 每日15:00 来源: 抖音热搜/小红书爆款/微博热搜 范围: 全国范围主动发现
拆解框架:
案例: [标题]
├─ 📊 数据快照 (点赞/播放/互动)
├─ 🔍 策略分析 (角度/情绪/格式)
├─ 🎯 可复制性评估 (置信度评分)
└─ 💡 电影小镇适配建议 (具体行动)
近期案例:
- 李白西湖对诗 → 356万播放,NPC随机对诗模式 → 可复制
- 重渡沟瀑布玫瑰裙 → 2.2万赞,场景改造+话题矩阵
- 万岁山520景区领证 → 前26对赠年卡 → 建议七夕档落地
- 韩国品牌抄袭汉服 → 微博第15位 → 国潮汉服日策划
时间: 每日16:00
方法: 轮换制 + 断点续做(/tmp/daily_task_state.json)
四平台采集清单:
| 平台 | 采集数据 | 工具 |
|---|---|---|
| 🎵 抖音指数 | 搜索指数/综合指数/关联词TOP10/人群画像 | cdp_keyword_deep.py |
| 📕 小红书灵犀 | 搜索量/热搜词/上下游词 | CDP + JS注入 |
| 📕 小红书搜索页 | 爆款笔记TOP10/标签/热搜问题 | CDP浏览器 |
| 🌐 百度搜索 | 营收/客流量/票价/媒体报道 | CDP浏览器 |
当前进度:已完成14/21个核心景区(截至5月25日)
时间: 周日10:00
系统每周计算自身预测准确率:
| 周次 | 准确率 | 趋势 |
|---|---|---|
| W18 | 60% | 基准 |
| W19 | 65% | +5pp |
| W20 | 70% | +5pp |
| W21 | 80% | +10pp |
W21错误模式分析:
- 执行层脱节:2次(建议已出但执行未跟上)
- 数据源断裂:1次(CSV断档8天)
- 信号误判:0次(持续改善)
| 来源 | 格式 | 位置 | 更新频率 |
|---|---|---|---|
| 主要:每日客流表 | 2026游客量统计.csv(宽表,368列) |
~/Desktop/ |
不定期(内部部门) |
| 历史参考 | 2023-2025年门票销售及客流统计数据表.xlsx |
~/Desktop/ |
年度 |
| 降级:飞书多维表格 | 电影小镇-2026年数量统计 |
飞书多维表 | 每日(CSV断档时) |
CSV是复杂的宽表格式(非标准的行-天结构),解析脚本 sync_obsidian_daily.py 处理方式:
行布局:
Row 0: 2023年参考 - 368个日值(用于同比)
Row 1: 2024年参考 - 368个日值
Row 2: 2025年参考 - 32个值(部分年份)
Row 3: 天气备注 - 每日天气
Rows 4-11: 门票细分 - 各渠道拆分
Row 12: 门票人数合计 - 主要指标
Row 13: 门票收入金额 - 门票收入(元)
Row 14: 闸机入园人次 - 实际入园
Rows 16-23: 预定数据 - 晨更/夕更
Row 25-28: 穿越德化街 - 演出数据
列映射:列索引2=1月1日,列3=1月2日,以此类推(从1月1日起的连续天)。
检测到CSV文件 → Python解析宽表
→ 提取:门票合计(Row12)、闸机入园(Row14)、天气(Row3)
→ 构建:{日期: {门票, 闸机, 天气}} 字典
→ 计算:日均、周环比、同比、YTD累计
→ 对比:年度目标(123万)、月度基准
→ 输出:结构化数据 → 飞书卡片(周报)
→ 记忆追加(每日)
→ Obsidian Wiki(数据.md更新)
~/Desktop/2026游客量统计.csv
│
├──→ scripts/sync_obsidian_daily.py(自动检测变更)
│ │
│ ├──→ workspace wiki/电影小镇/历史数据/2026年/数据.md
│ │ (自动更新时间戳)
│ │
│ └──→ Obsidian Vault(同步目标)
│ (镜像workspace wiki至用户Obsidian)
│
└──→ Agent周度客流洞察任务(周二9:30)
读取CSV → 生成5章卡片
→ 推送至飞书群
→ 记录至memory
| 指标 | 计算方式 | 用途 |
|---|---|---|
| 日总量 | Row 12, 第N列 | 基础数据点 |
| 月日均 | 月度合计 ÷ 天数 | 容量规划 |
| YTD完成率 | 累计 ÷ 1,230,000 | 目标追踪 |
| 周环比 | (本周-上周)÷上周 | 趋势检测 |
| 同比 | (2026-2023/2024)÷基准 | 增长分析 |
| 渠道占比 | Rows 4-11 ÷ 合计 | 渠道效率 |
| 闸机转化率 | Row 14 ÷ Row 12 | 核销率 |
| 指标 | 数值 |
|---|---|
| CSV最后更新 | 2026-05-17 |
| 数据断档 | 8天( |
| YTD合计 | 656,067 |
| 目标完成率 | 42.9% |
| 有数据天数 | 135天(1月1日-5月17日) |
| 单日最高 | 33,411(5月3日,五一) |
穿越德化街是建业电影小镇的旗舰室内剧场演出--一场设定在1930年代郑州德化街的穿越沉浸式大秀。剧场于2024年底完成扩建:
- 扩建前(2023): 每场450座
- 扩建后(2025+): 每场1,140座(+253%)
- 扩建影响: 2024年10-11月闭园施工,12月仅压力测试
演出数据同样存储在客流CSV中(Rows 25-28):
| CSV行 | 数据 | 单位 |
|---|---|---|
| Row 25 | 日期标签 | 字符串("1月1日") |
| Row 26 | 场次 | 计数 |
| Row 27 | 库存(座位总量) | 座位数(每场1,140) |
| Row 28 | 售卖(售出门票) | 张数 |
| (派生) | 上座率 | 售卖÷库存 |
1. 演出节奏
- 平日:1场/天(1,140座)
- 高峰周末:2-3场/天
- 节假日峰值:5-6场/天(2026五一5天23场)
2. 上座率追踪
上座率 = 售出门票 ÷ (场次 × 1,140)
阈值:
🔴 低于40% → 表现不佳,需检查定价/内容
🟡 40-65% → 正常范围
🟢 65-85% → 良好利用
💎 高于85% → 容量瓶颈,考虑加场
3. 转化率(核心指标)
转化率 = 观演人次 ÷ 入园人次
历史数据:
2023: 18.0%(基准,扩建前)
2024: 16.2%(下降,扩建前)
2025: 35.2% 🚀(翻倍,扩建后--产品力提升)
2026 YTD: 27.0%(追踪中--Q1淡季低频)
4. 票种结构(套票 vs 加购)
| 指标 | 2023 | 2024 | 2025 | 2026 Q1 |
|---|---|---|---|---|
| 套票占比 | 72.5% | 60.7% | 77.6% | 49.5% |
| 加购占比 | 27.5% | 39.3% | 22.4% | 50.5% |
| 套票单价 | - | - | ¥101.59 | - |
| 加购单价 | - | - | ¥52.80 | - |
| 月份 | 天数 | 场次 | 观演人次 |
|---|---|---|---|
| 4月 | 30 | 52 | 37,581 |
| 5月(至17日) | 17 | 47 | 33,565 |
| Q2合计 | 47 | 99 | 71,146 |
五一峰值(5月1日-4日): 5-6场/天,92-93%上座率--接近容量上限。
CSV Rows 25-28 解析 → sync_obsidian_daily.py
→ 提取:日期、场次、库存、售卖
→ 计算:上座率、转化率、趋势
→ 更新:
wiki/电影小镇/演出节目/穿越德化街.md
wiki/sources/穿越德化街数据分析.md
→ 周报:纳入客流洞察卡片
| 来源 | 数据 | 更新 |
|---|---|---|
| 客流CSV Row 13 | 门票收入金额 | 随CSV同步 |
| 2023-2025年门票销售及客流统计表.xlsx | 历史营收 | 年度 |
| 穿越德化街Excel | 演出收入 | 按需获取 |
总营收 = 门票收入 + 非门票收入
│ │
▼ ▼
× 日客流 餐饮、购物、住宿、
× 平均票价 旅拍、体验项目
门票收入渠道:
| 渠道 | 说明 | 数据源 |
|---|---|---|
| 线上散客 | 小程序/抖音/美团 | CSV Rows 9-10 |
| 窗口散客 | 现场购票 | CSV Row 10 |
| 旅行社 | 团队游 | CSV Row 8 |
| 大客户 | 企业活动 | CSV Row 7 |
| 研学 | 教育团体 | CSV Row 6 |
| 年份 | 演出收入 | 说明 |
|---|---|---|
| 2023 | ¥3,139万 | 扩建前基准 |
| 2024 | ¥2,675万 | 扩建前,-14.8% |
| 2025 | ¥4,266万 | 🚀 扩建后,+35.9% |
| 2026 Q1 | ¥632万 | 追踪中 |
| 分析 | 方法 | 输出 |
|---|---|---|
| 平均票价(ATP) | 收入÷客流 | 定价策略输入 |
| 渠道占比 | 各渠道÷合计 | 渠道优化 |
| 非门票收入占比 | (总-门票)÷总 | 辅助收入追踪 |
| 演出收入贡献 | 演出收入÷总营收 | 产品效率 |
| 维度 | 取值 | 说明 |
|---|---|---|
| 🎯 影响等级 | 🔴高/🟡中/🟢低/📡噪音 | 对电影小镇的实际影响程度 |
| 💡 建议动作 | 跟风/借势/警惕/忽略 | 团队应该做什么 |
| ⏰ 执行窗口 | 今天/本周/不紧急 | 必须在什么时间前行动 |
| 明确的损失陈述 | 不做的后果量化 |
D<序号> | <一句话标题>
├─ 简单理解:一句话解释
├─ 做错风险:不执行的后果
├─ 推荐+理由:具体建议 + 置信度评分
└─ 利弊:收益 vs 成本/精力
| ❌ 禁止 | ✅ 替换为 |
|---|---|
| "值得关注" | "本周必须执行" 或 具体优先级 |
| "仅供参考" | 具体建议+置信度 |
| "或许可以考虑" | "推荐:..." |
| "需要进一步分析" | 明确判断 或 明确"信息不足" |
KNOWLEDGE LAYER (wiki/)
│
├── concepts/ ← 抽象模式与理论
│ ├── 演艺景区.md - 演艺景区特征
│ ├── 内容爆款规律.md - 爆款内容模式
│ ├── 景区营销漏斗.md - 漏斗模型
│ ├── 情绪营销.md - 情绪营销框架
│ ├── 平台算法规则.md - 抖音/小红书算法
│ └── ... (共12个概念文件)
│
├── entities/ ← 具体对象
│ ├── 建业电影小镇.md - 自身档案
│ ├── 万岁山武侠城.md - 竞品
│ ├── 清明上河园.md - 竞品
│ ├── 抖音平台.md - 平台实体
│ ├── 小红书平台.md - 平台实体
│ └── ... (共12个实体文件)
│
├── sources/ ← 数据溯源
│ ├── 穿越德化街数据分析.md
│ ├── 抖音指数追踪日报.md
│ ├── 竞品深度档案.md
│ └── ... (共8个来源文件)
│
└── queries/ ← 决策问答归档
├── 抖音与小红书平台差异.md
├── 知识层与业务层关系.md
└── ... (共5个问答文件)
BUSINESS LAYER (wiki/电影小镇/)
├── 基础档案.md - 基本信息/年度目标(123万/1.2亿)
├── 战略框架.md - SWOT/竞争定位
├── 人群画像.md - 抖音/小红书用户画像
├── 历史数据/ - 2023-2026历年客流
│ ├── 数据.md - ✅ YTD 656,067 / 42.9%
│ ├── 规律洞察.md - 春节/暑期/国庆三峰值
│ └── 2023年/2024年/2025年/
├── 演出节目/
│ └── 穿越德化街.md - 6年数据 + Q2更新
│ (240场/177,076人次/转化率27.0%)
├── 运营方法/ - 抖音/小红书运营SOP
└── 运营规划/ - 夏季运营方案
COMPETITOR INTELLIGENCE (wiki/竞品分析/)
├── 竞品深度档案/ - 20个全国景区深度分析
│ ├── 郑州方特欢乐世界深度分析.md
│ ├── 银基动物王国深度分析.md
│ ├── 清明上河园深度分析.md
│ ├── 大唐不夜城深度分析.md
│ ├── 阿那亚深度分析.md
│ └── ... (20个文件,全部完成)
│
├── 竞品动态追踪/ - 每日动态日志 (4月20-27日)
├── 追踪数据/
│ ├── 抖音指数追踪.md
│ └── 小红书爆款追踪.md
│
└── 关键词池状态.md - 断点续做检查点
CASE LIBRARY (wiki/全国景区案例库/)
├── index.md - 10条爆款公式索引
└── 20+ 按周归档案例
├── 大唐不夜城夜游标杆-2026W17.md
├── 万岁山武侠城标杆-2026W17.md
├── 打铁花跨景区爆款现象-2026W17.md
├── 乌镇住宿早茶客46%复购率-2026W22.md
└── ...
10条爆款公式(W22更新):
| # | 公式 | 案例 |
|---|---|---|
| 1 | 情绪悬念+反转曝光 | 万岁山/大唐不夜城 |
| 2 | 非遗x实景融合 | 打铁花/大宋演出 |
| 3 | 本地KOC矩阵分发 | 多角度覆盖策略 |
| 4 | NPC随机惊喜互动 | 对诗/领证/情景剧 |
| 5 | 季节限定场景改造 | 夏日夜市/樱花季 |
| 6 | 反向操作·安静≠无聊 | 只有河南"不躁"音乐会 |
| 7 | 情绪经济x情感绑定 | 初雪许愿/日落摄影 |
| 8 | 空间文化再造 | 瀑布玫瑰裙/悬崖咖啡 |
| 9 | 节庆IP长期运营 | 清明系列/端午 |
| 10 | 用户共创UGC话题 | "我的隐藏打卡点" |
| 选项 | 评估 | 结论 |
|---|---|---|
| OpenClaw + Agent Skills | ✅ 选择 | 25+模型供应商、cron调度、技能生态、本地优先 |
| 纯Python脚本 | ❌ 放弃 | 无Agent推理能力,无法自主决策 |
| 付费SaaS | ❌ 放弃 | 昂贵,无法定制决策框架,数据锁定 |
微信小程序不支持npm d3-force(ES模块+DOM依赖),图谱可视化使用手写Verlet积分物理引擎:
const kRepulsion = 700; // 库仑排斥力
const kSpring = 0.04; // 胡克弹簧张力
const lLength = 120; // 理想弹簧长度
const gravity = 0.02; // 中心引力
const friction = 0.85; // 阻尼衰减性能:50个节点下稳定60fps。
抖音创作者平台使用重度客户端渲染(SPA),Playwright脚本有时返回空数据或截断数据。CDP浏览器直连作为可靠降级通道:
Script → [成功?] → 解析使用
↓
[失败?] → CDP直连提取 → 解析使用
↓
[失败?] → 基于近期数据推断趋势
飞书默认message tool发送msg_type: "post"(富文本),不渲染表格。card脚本发送msg_type: "interactive" + schema: "2.0":
payload = {
"receive_id": chat_id,
"msg_type": "interactive",
"content": json.dumps(card, ensure_ascii=False)
}除核心模块外，系统还包含丰富的支持工具、运维机制和质量保障流程。
| 脚本 | 用途 | 类型 | 特点 |
|---|---|---|---|
douyin_index.py |
抖音指数采集 | 📊 数据 | 双通道(Playwright+CDP) |
competitor_keyword_v8.py |
竞品关键词深度分析 | 📊 数据 | 四平台聚合 |
cdp_cookie_hub.py |
CDP浏览器Cookie提取 | 🔧 运维 | 跨平台(抖音/小红书/微博) |
cdp_keyword_deep.py |
CDP深度关键词分析 | 📊 数据 | Tab2/4自动切换 |
cdp_collect.py |
通用CDP数据采集 | 📊 数据 | Tab感知定位 |
xiaohongshu_crawl.py |
小红书数据采集 | 📊 数据 | 反限流处理 |
xhs_competitor_crawl.py |
小红书竞品页面抓取 | 📊 数据 | 多关键词队列 |
send_feishu_card.py |
飞书交互卡片发送 | 📨 推送 | schema 2.0验证 |
sync_obsidian_daily.py |
Obsidian Wiki同步 | 🔄 同步 | CSV→Wiki→Vault流水线 |
query_passenger.py |
客流CSV查询助手 | 📊 数据 | 日期范围筛选 |
self_check.py |
系统自检诊断 | 🔧 运维 | 8项健康检查清单 |
cdp_restore_tabs.py |
CDP浏览器Tab恢复 | 🔧 运维 | 崩溃自动恢复 |
case_library_scan.py |
案例库审计修复 | 🔄 同步 | 断链检测 |
industry_news_browser.py |
文旅情报新闻采集 | 📊 数据 | 多源聚合 |
llmwiki_ingest.py |
Wiki知识接入 | 🔄 同步 | karpathy-wiki模式 |
validate_data.py |
数据完整性验证 | ✅ 质控 | 过时&完整性检查 |
wiki_drift_check.py |
Wiki漂移检测 | ✅ 质控 | 交叉引用验证 |
总计：45+个活跃脚本（不含已归档/遗留脚本）
所有报告以飞书交互卡片（interactive card）形式推送，非纯文本。卡片系统拥有独立的设计规范、验证脚本和故障恢复机制。
{
"schema": "2.0", // 必须
"header": {
"title": {
"tag": "plain_text", // 非"lark_md"
"content": "📊 抖音指数日报 | 2026-05-25"
},
"template": "blue"
},
"body": {
"elements": [
{
"tag": "markdown",
"content": "| 景区 | 搜索指数 | ... |"
}
]
}
}| 规则 | 原因 | 建立时间 |
|---|---|---|
header.title.tag必须为"plain_text" |
"lark_md"导致渲染异常 |
2026-04-19 |
内容放在body.elements[]，非根级elements[] |
Schema合规 | 2026-04-19 |
表格用|管道符放在markdown元素内 |
飞书MD解析器 | 2026-04-19 |
换行用<br/>不用\n |
飞书忽略\n |
2026-04-10 |
必须用send_feishu_card.py，非默认message tool |
msg_type:"post"不渲染表格 |
2026-05-25 |
章节分隔符：━━━━━━━━━━━━━━━━━━━━━━━━
Emoji层级： 📌 → 🔍 → ⚠️ → 💡
加粗强调： **文字**
等宽代码： `等宽文字`
决策输出： D<序号> | <一行标题>
# 1. 验证卡片结构
validate_card(card) → 检查schema/header/elements
# 2. 获取飞书tenant token
token = get_token() # client_credentials授权
# 3. 以正确msg_type发送
payload = {
"receive_id": chat_id,
"msg_type": "interactive", # 非"post"
"content": json.dumps(card, ensure_ascii=False)
}
# 4. Token过期自动重试
if result.code in (99991663, 19001):
token = refresh_token()
retry()| 日期 | 故障 | 根因 | 修复 |
|---|---|---|---|
| 05-25 | W22营销日历显示为纯文本 | Agent用了默认message tool(msg_type:"post") | 9个cron任务统一为send_feishu_card.py |
| 04-27 | 卡片表格不渲染 | header.title.tag为"lark_md" |
改为"plain_text" |
| 04-19 | elements在根级而非body.elements[] |
Agent误解schema | 在send_feishu_card.py中增加验证 |
系统使用多层降级策略保障每日运行不间断：
第一层：主脚本
└─ douyin_index.py运行 → [成功?]
├─ 是 → 解析+使用
└─ 否 → 降级至第二层
第二层：CDP直连提取
└─ CDP浏览器导航至目标页 → [成功?]
├─ 是 → 解析+使用
└─ 否 → 降级至第三层
第三层：趋势推断
└─ 基于近期数据+已知模式估算
→ 标记为"估算"（低置信度）
| 错误模式 | 触发条件 | 响应 | 恢复方式 |
|---|---|---|---|
| 503服务繁忙 | DeepSeek API过载 | 下一cron周期自动重试 | 无需操作（服务商） |
| Cookie过期 | 抖音/小红书登录超时 | CDP浏览器降级；下一同步周期 | 扫码重新登录 |
| 667字符截断 | 抖音SPA页面部分渲染 | 自动切换CDP提取 | 等待脚本更新 |
| 工具执行超时 | CDP浏览器繁忙/负载高 | Agent以30s超时重试 | 监控，无需手修 |
| 编辑失败 | Wiki文件写入冲突 | Agent重试新内容 | 极少发生，自动恢复 |
cron任务的告警配置示例：
{
"failureAlert": {
"after": 2, // 连续2次失败后告警
"channel": "feishu",
"to": "oc_f109bcfd1bc7e166fd0ae077f70247cf",
"cooldownMs": 60000 // 告警间隔1小时
}
}Agent的能力通过OpenClaw的Skills系统可扩展。技能来源：ClawHub（社区仓库）+ 自定义编写。
| 技能 | 用途 | 类型 |
|---|---|---|
wechat-mini-program-builder |
微信小程序快速搭建 | 📱 开发 |
mini-program-dev |
小程序代码模板&API | 📱 开发 |
wechat-miniprogram-skill |
小程序从入门到精通指南 | 📱 开发 |
miniprogram-development |
通用小程序开发 | 📱 开发 |
frontend-design-3 |
前端设计规范 | 🎨 UI |
react-best-practices |
React最佳实践 | 💻 代码 |
typescript-skills |
TypeScript技能集 | 💻 代码 |
karpathy-coding-guidelines |
Karpathy编码准则 | 💻 代码 |
debug-pro |
系统化调试 | 🔧 开发 |
api-tester |
HTTP请求测试(GET/POST/PUT/DELETE) | 🔧 开发 |
browser-automation |
自然语言浏览器自动化 | 🤖 自动化 |
karpathy-guidelines |
LLM编程通用智慧 | 💻 代码 |
可用技能总数：59（系统内置 + 工作区 + 已安装）
Agent维护两种记忆：情景记忆（日常日志）和语义记忆（精炼规则），以及反馈闭环实现持续改进。
情景记忆（原始日志）
memory/
├── YYYY-MM-DD.md ← 每日执行记录
语义记忆（精炼）
MEMORY.md ← 长期规则（最长100行）
├── 铁律（不可违反）
├── 关键洞察（可复用模式）
├── [reference]（系统指针）
├── [project]（项目状态）
└── [feedback]（纠错/确认）
状态（机器可读）
memory/heartbeat-state.json
memory/topics/
├── feedback/ ← 用户纠错&确认
└── projects/ ← 长运行项目状态
用户说"不要"/"不对"/"停止"
→ Agent记录纠错至memory/topics/feedback/
→ 如模式确认则更新MEMORY.md规则
→ 调整后续行为
用户说"对"/"很好"/"就这样"
→ Agent记录确认（成功模式）
→ 强化已有规则
→ 无需变更
| 类型 | 用途 | 示例 |
|---|---|---|
[user] |
用户角色/偏好/目标 | user: 站长偏好详细数据报告 |
[feedback] |
工作指导（纠错+确认） | feedback: 达人必须绑定转化链路 |
[project] |
项目状态/目标 | project: 当前在优化多Agent系统 |
[reference] |
外部系统指针 | reference: 飞书群 oc_xxx |
检查清单：
□ cron任务正常运行（检查lastError）
□ CDP浏览器在线（端口18800）
□ Cookie文件未过期
□ 客流CSV未过时（更新<14天）
□ 飞书Bot Token有效
□ 磁盘空间充足（<90%）
□ 无orphan会话文件（>100需清理）
□ Skills&Plugins加载无错误
| 任务 | 时间 | 说明 |
|---|---|---|
| Wiki健康检查 | 10:00 | 验证所有Wiki链接，修复断链 |
| 代码漂移检查 | 10:00 | 检测workspace与Wiki内容漂移 |
| Orphan会话清理 | 11:00 | 归档/删除未使用的.jsonl文件 |
| 技能探索 | 14:00 | 发现ClawHub新技能，更新技能集 |
| 周度演进评审 | 每周 | 系统升级、记忆整理、SOP审计 |
| 属性 | 详情 |
|---|---|
| 端口 | 18800 |
| 目标 | host（Mac Mini） |
| Tab 0 | 小红书灵犀后台 |
| Tab 1 | 百度搜索 |
| Tab 2 | 抖音订阅页 |
| Tab 3 | 抖音iframe |
| Tab 4 | 抖音关键词页 |
| Tab 5 | 抖音iframe |
| Tab 6 | 小红书探索页 |
| Cookie存储 | /tmp/juLiang_cookies.json（抖音） |
/tmp/xiaohongshu_cookies.json（小红书） |
|
/tmp/weibo_cookies.json（微博） |
|
| 代理 | 127.0.0.1:7897（抖音专用） |
| 恢复 | cdp_restore_tabs.py Tab崩溃恢复 |
| 门控 | 检查项 | 失败处理 |
|---|---|---|
| ✅ | 脚本返回全部8个景区 | 部分数据→切换CDP |
| ✅ | Cookie时效<24h | 过期→CDP降级，标记需刷新 |
| ✅ | CSV更新<2周 | 过期→周报标记，请求同步 |
| ✅ | 飞书卡片送达确认 | 失败→Token刷新重试 |
| ✅ | Wiki写入成功 | 冲突→唯一内容重试 |
每周日，系统执行结构化演进周期：
| 阶段 | 行动 | 产出 |
|---|---|---|
| 🧹 清理 | 归档orphan会话、压缩记忆 | 干净状态 |
| 📚 接入 | 处理新Wiki内容（raw/→知识层） | 更新Wiki |
| 🔍 审计 | 检查预测准确率、错误模式 | 准确率报告 |
| 🧠 学习 | 基于验证结果更新决策规则 | 规则更新 |
| 🚀 探索 | 搜索ClawHub新技能 | 技能更新 |
| 📝 提交 | Git commit & push Wiki更新 | GitHub同步 |
| 资源 | 用量 | 管理策略 |
|---|---|---|
| API Token | ~20万token/天 | 单一模型最大化上下文窗口；isolated session防止状态膨胀 |
| 磁盘 | 81Gi可用/228Gi总 | 每周orphan清理；已删除Ollama模型（回收26Gi） |
| CDP浏览器 | 6个永久Tab | Tab级精确定位防止资源浪费 |
| Cron会话 | 23个isolated session | 临时性（运行后删除）；每任务30s-600s超时 |
| 飞书API | ~30次调用/天 | Token缓存减少认证请求 |
| 指标 | 数值 |
|---|---|
| 连续运行天数 | 30+天 |
| 日均产出决策简报 | 6-8份 |
| 日均消耗API token | ~20万(DeepSeek-V4-Flash) |
| 已推送飞书卡片 | 180+张 |
| 活跃cron任务 | 23个 |
| 允许模型 | deepseek/deepseek-v4-flash(唯一) |
| 资产类型 | 数量 | 说明 |
|---|---|---|
| 竞品深度档案 | 20个 | 四平台全量分析 |
| 爆款案例 | 20+个 | 每周更新 |
| 爆款公式 | 10条 | 已验证的模式 |
| SOP规范 | 22个 | 标准化操作手册 |
| 历史数据 | 4年 | 2023-2026年客流 |
| Wiki文件 | 150+篇 | Markdown文档 |
- 搜索指数回升:电影小镇搜索指数W21环比**+24.51%**(品牌修复信号)
- 竞品排名:从第7升至第4位
- 预测准确率:从W18的60%持续提升至W21的80%
- 内容真空检测:成功识别"搜索涨·综合跌"背离模式(5月22日验证)
| 层级 | 技术 | 用途 |
|---|---|---|
| AI框架 | OpenClaw v2026.5.20 | Agent运行时/cron/工具 |
| 大模型 | DeepSeek-V4-Flash | 推理/分析/生成 |
| 数据采集 | Python 3.12 + Playwright + CDP | 网页抓取/浏览器自动化 |
| 推送渠道 | 飞书Bot API (interactive card) | 报告推送至群聊 |
| 知识库 | Obsidian + karpathy-wiki | 三层抽象结构化Wiki |
| 操作系统 | macOS 26.4 (Darwin 25.4.0) ARM64 | 宿主机环境 |
| 硬件 | Mac Mini (2024) | 24/7本地运行 |
| Node.js | v25.8.2 | OpenClaw运行时 |
| 技能仓库 | ClawHub (13,729个可用技能) | 可扩展的Agent能力 |
| 版本控制 | Git + GitHub | 代码与Wiki仓库 |
在Obsidian中打开 wiki/ 目录。从 wiki/index.md 开始浏览。
所有报告自动推送至飞书群。执行摘要记录在 memory/YYYY-MM-DD.md。
scripts/- Python自动化脚本(45+文件)wiki/- 知识库(150+ Markdown文件)
系统包含反馈闭环。纠错记录在 memory/topics/feedback/ 并影响后续行为。
系统正在持续演进中，以下功能已在规划或开发中：
| 功能 | 状态 | 说明 |
|---|---|---|
| 🔴 爆款视频自动拆解 | 🟡 规划中 | 自动爬取抖音热门视频 → AI分析 → 可复制性评估 → 飞书卡片。当前为手动选择，目标：全自动每日流水线。 |
| 🔴 客流实时看板 | 🟡 规划中 | 替换stale CSV依赖，接入园区票务系统实时API。自动检测异常（日偏差>20%）。 |
| 🟣 营收追踪模块 | 🔴 研究中 | 整合票务+演出系统营收数据。自动ATP计算、营收结构分析、非门票收入追踪。 |
| 🟣 竞品定价监控 | 🔴 研究中 | 每日自动检查竞品票价（抖音团购/美团/官方小程序）。价格变动>10%即时告警。 |
| 功能 | 状态 | 说明 |
|---|---|---|
| 🟢 周度预测准确率 | ✅ 已运行 | 已实现（W21: 80%）。通过学习闭环持续改进。 |
| 🟣 UGC情感分析 | 🔴 研究中 | 自动分析小红书/抖音评论情感趋势。负面情绪飙升早期预警。 |
| 🟣 自动内容生成 | 🔴 研究中 | 基于热门格式生成内容草稿（短视频脚本/小红书笔记）。人工审核后发布。 |
| 🟣 多Agent协作 | 🟡 规划中 | 专业化子Agent：竞品Agent/内容Agent/客流Agent/复盘Agent——在协调者下并行工作。 |
| 功能 | 状态 | 说明 |
|---|---|---|
| 🟣 公共API层 | 🔴 研究中 | 对外暴露匿名化竞品情报数据API，供其他景区参考。 |
| 🟣 模板市场 | 💡 构想 | 可分享的SOP模板、决策框架、报告卡片模板。 |
| 🟣 跨场景对标 | 💡 构想 | 电影小镇指标与全国均值对比（季节性调整）。 |
| 🟣 微信小程序延展 | 🟡 开发中 | ChatWiki小程序（全息知识图谱 + AI析构中心）—— 开发进行中。 |
| 图标 | 含义 |
|---|---|
| 🟢 | 已上线 — 正在运行 |
| 🟡 | 规划中 — 设计/方案进行中 |
| 🔴 | 研究中 — 可行性验证 |
| 💡 | 构想 — 概念阶段 |
由AI构建，服务于人。2026年4月启动运行。
由OpenClaw AI Agent自动部署与维护。
最后系统更新：2026-05-25