AI 前沿情报汇总

> 信息聚合 ≠ 信息堆砌。每日情报经筛选、评分、分层后输出，不做 50 条标题的噪音。

When to Use

触发词（任意语言）：

• "AI 前沿" / "情报汇总" / "每日情报" / "frontier monitor" / "daily briefing" → 全量简报
• "今天有什么信号" / "signal check" / "快速扫描" / "what signals today" → 快速信号检测
• "arXiv 最新" / "论文追踪" / "paper tracker" / "latest papers" → 仅 arXiv 轨道
• "GitHub Trending" / "AI 热榜" / "trending AI" → 仅 GitHub 轨道

Architecture: 5-Track Parallel

> For full data source details, read {baseDir}/references/data-sources.md

Workflow

Step 1: Fetch All Tracks

# Track 1: RSS (run crawler first, outputs to {baseDir}/data/candidates/)
python3 {baseDir}/scripts/rss-crawler.py

# Track 2-4: Generate briefing (all tracks auto-fetched)
python3 {baseDir}/scripts/generate-briefing.py --mode full

Modes: full | quick | arxiv | github

Step 2: Auto-Score & Tier

Each candidate without a score is auto-scored (0-5) by keyword matching across 4 dimensions:

Source bonus: OpenAI/AWS +1.0, Techmeme +0.5, PH/HN +0.3

Tiers are dynamic (based on actual score distribution, not hardcoded thresholds):

• 🔴 Core: top ~15% or ≥3.5 (max 3)
• 🟡 Worth watching: top ~30% or ≥2.5 (max 5)
• 🟢 Quick scan: ≥1.0 (max 8, 36kr first)

> For scoring keywords and signal detection rules, read {baseDir}/references/scoring.md

Step 3: Detect Signals

Extract cross-track signals into 3 dimensions:

• 🛠 Tech trends — new models, architectures, frameworks, benchmarks
• 🏢 Product launches — new releases, open-source, GA announcements
• 💰 Funding/M&A — investments, acquisitions, IPOs

Step 4: Render Briefing

Strict format — emoji headers, tiered sections, signal summary. Output in Chinese (中文为主). Total ≤ 16 items across all tiers.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🤖 AI 前沿情报 · {Day} {Date}
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📡 数据源：11 RSS + 36kr + arXiv + GitHub + Anthropic
   候选：{N} 条 | 高质量：{M} 条 | 阈值：核心≥{X} / 关注≥{Y}

## 🔴 核心情报（{N} 条）
### 1. {Title}
🔗 {Link}
💡 启示：{One-line insight}

## 🟡 值得关注（{N} 条）
1. **{Title}**
   🔗 {Link}

## 🟢 快速浏览（{N} 条）
• [{Title}]({Link})

## 📚 arXiv · 论文追踪（≤3 篇）
**{Title}** — {Authors} | {Date}
摘要：{Abstract[:150]}... → {Link}

## 🔥 GitHub Trending · AI（≤3 个）
**{Repo}** ({Lang}) +{TodayStars}⭐ → {Link}

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📊 今日信号
🛠 技术趋势：{signal}
🏢 产品发布：{signal}
💰 资本动向：{signal}
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⏰ {HH:MM} | ai-frontier-monitor v3.0

Step 5: Deliver & Archive

1. Reply in conversation — 直接在当前对话输出简报
2. Push to Feishu — 通过 message 工具发送到飞书（channel: feishu, to: user ID）
3. Save to file — 将完整简报保存为 Markdown 文件到：

```

{baseDir}/data/briefings/{YYYY-MM-DD}-frontier-briefing.md

```

保存时覆盖当日内容。

Data Directory

All runtime data is stored under {baseDir}/data/:

{baseDir}/data/
├── candidates/          # RSS 爬取的候选条目 (JSON)
│   └── *_candidates.json
├── briefings/           # 生成的简报 (Markdown)
│   └── YYYY-MM-DD-frontier-briefing.md
└── rss-state.json       # RSS 爬取状态

> {baseDir} is the skill root directory containing this SKILL.md. All paths use {baseDir} for portability.

Edge Cases

Skill Integration

Configuration

Edit {baseDir}/references/BRIEFING_CONFIG.md to customize:

• Quantity limits per tier
• Data source on/off switches
• Signal detection thresholds
• Delivery target (Feishu user ID / Discord channel / etc.)

Quality Gates

• Max 16 items per day (3+5+5+3 papers)
• NO_REPLY when <2 quality candidates
• Deduplicate same company/product, keep highest score
• 3 consecutive days below 3 core items → trigger keyword review

Dependencies

• Python 3.8+ with feedparser (for RSS crawling)
• bash (for arXiv/GitHub fetch scripts)
• curl (for 36kr API)
• web_search tool (for Anthropic track)

_Last updated: 2026-05-09 | v3.0_