AI Daily Digest

Fetches the latest articles from 90 popular tech blogs recommended by Karpathy, then the Agent (you) will score them, filter by quality, and generate a curated daily digest with summaries and trend analysis.

Command

`/digest`

Run the daily digest generator.

Usage: Type /digest, and follow the interactive guide to collect parameters, then the Agent will process everything.

How It Works

Two-Phase Architecture:

Phase 1 - Script (Fetch Only): digest.ts fetches RSS feeds, filters by time, outputs raw JSON
Phase 2 - Agent (Intelligence): You (the Agent) score articles, summarize them, generate trends, build the Markdown report

No external API keys needed — everything uses the Agent's own intelligence.

Script Directory

All scripts are in scripts/ subdirectory.

Script	Purpose
`scripts/digest.ts`	Fetch-only script - RSS fetching + time filtering → JSON output

Configuration Persistence

Configuration file path: ~/.tomtom-daily-digest/config.json

Agent must check this file before execution:

If exists, ask user whether to reuse saved config
After execution, save current config to this file

Configuration structure:

{
  "timeRange": 48,
  "topN": 15,
  "language": "zh-tw",
  "lastUsed": "2026-02-14T12:00:00Z"
}

Note: Language options are zh-tw (Traditional Chinese) or en (English) only.

Interactive Flow

Usage Notice

Output this notice at the beginning every time /digest is run:

Step 0: Check Saved Configuration

cat ~/.tomtom-daily-digest/config.json 2>/dev/null || echo "NO_CONFIG"

If configuration exists, ask whether to reuse:

question({
  questions: [{
    header: "Use saved config",
    question: "A previously used configuration was detected:\n\n• Time range: ${config.timeRange} hours\n• Selection count: ${config.topN} articles\n• Output language: ${config.language === 'zh-tw' ? '繁體中文' : 'English'}\n\nPlease choose an action:",
    options: [
      { label: "Use last config and run (Recommended)", description: "Use all saved parameters and start immediately" },
      { label: "Reconfigure", description: "Configure all parameters from scratch" }
    ]
  }]
})

Step 1: Collect Parameters

Use question() to collect all at once:

question({
  questions: [
    {
      header: "Time range",
      question: "How far back should articles be fetched?",
      options: [
        { label: "24 hours", description: "Last day only" },
        { label: "48 hours (Recommended)", description: "Last two days, broader coverage" },
        { label: "72 hours", description: "Last three days" },
        { label: "7 days", description: "Articles from the past week" }
      ]
    },
    {
      header: "Selection count",
      question: "How many articles to keep after AI filtering?",
      options: [
        { label: "10 articles", description: "Compact version" },
        { label: "15 articles (Recommended)", description: "Standard recommendation" },
        { label: "20 articles", description: "Extended version" }
      ]
    },
    {
      header: "Output language",
      question: "What language should the digest be in?",
      options: [
        { label: "繁體中文 (Recommended)", description: "Summaries translated to Traditional Chinese" },
        { label: "English", description: "Keep original English text" }
      ]
    }
  ]
})

Step 2: Execute Fetch Script

Run the script to fetch articles and save JSON to a temp file:

SKILL_DIR="<path to this SKILL.md's directory>"
TEMP_JSON="/tmp/digest-articles-$(date +%s).json"

npx -y bun ${SKILL_DIR}/scripts/digest.ts \
  --hours <timeRange> \
  --output "$TEMP_JSON"

The script will output JSON like:

{
  "metadata": {
    "totalFeeds": 90,
    "successfulFeeds": 85,
    "totalArticles": 450,
    "filteredArticles": 120,
    "timeRangeHours": 48,
    "fetchedAt": "2026-02-14T..."
  },
  "articles": [
    {
      "title": "...",
      "link": "...",
      "pubDate": "2026-02-14T...",
      "description": "...",
      "sourceName": "simonwillison.net",
      "sourceUrl": "https://..."
    }
  ]
}

Step 3: Read and Parse JSON

Use the Read tool to load the JSON:

Read($TEMP_JSON)

Parse the JSON and extract the articles array.

Step 4: Agent Scoring (You Do This)

IMPORTANT: You (the Agent) will now score all articles.

For each article, assign scores (1-10) on three dimensions:

Scoring Dimensions

1. Relevance - Value to tech/programming/AI/internet professionals

10: Major event/breakthrough everyone should know
7-9: Valuable to most tech practitioners
4-6: Valuable to specific tech domains
1-3: Low relevance to tech industry

2. Quality - Article depth and writing quality

10: Deep analysis, original insights, rich citations
7-9: Has depth, unique perspective
4-6: Accurate info, clear expression
1-3: Shallow or purely reposting

3. Timeliness - Whether worth reading now

10: Major ongoing event / just-released important tool
7-9: Related to recent hot topics
4-6: Evergreen content, not outdated
1-3: Outdated or no timeliness value

Category Assignment

Assign one category to each article:

ai-ml: AI, machine learning, LLM, deep learning
security: Security, privacy, vulnerabilities, encryption
engineering: Software engineering, architecture, programming languages, system design
tools: Dev tools, open source projects, newly released libraries/frameworks
opinion: Industry opinions, personal reflections, career development, cultural commentary
other: None of the above fit well

Keyword Extraction

Extract 2-4 keywords that best represent the article's topic (use English, short, like "Rust", "LLM", "database", "performance").

Processing Strategy

If >50 articles: Process in batches of 15-20 to avoid context overflow. Score batch 1, then batch 2, etc.

If ≤50 articles: Score all at once.

Output format (internal, keep in memory):

Article scores = [
  {
    index: 0,
    relevance: 8,
    quality: 7,
    timeliness: 9,
    category: "engineering",
    keywords: ["Rust", "compiler", "performance"],
    totalScore: 24
  },
  ...
]

Step 5: Select Top N Articles

Sort all scored articles by totalScore (sum of relevance + quality + timeliness) descending.

Take the top N articles (based on user's selection: 10, 15, or 20).

Step 6: Generate Summaries for Top Articles

For each of the top N articles, generate:

Translated Title (titleTranslated):
- If language is zh-tw: Translate the English title to natural Traditional Chinese (繁體中文)
- If language is en: Keep the original English title
- If original is already in the target language, keep unchanged
Summary (4-6 sentences): Structured summary that lets readers understand the core content without clicking. Include:
- Core problem or topic the article discusses (1 sentence)
- Key arguments, technical solutions, or findings (2-3 sentences)
- Conclusion or author's core viewpoint (1 sentence)
Recommendation Reason (reason): One sentence explaining "why worth reading", distinct from summary (summary says "what", reason says "why").

Language: Use the language specified by user (zh-tw or en).

Summary requirements:

Directly state key points, don't use openings like "This article discusses..." or "本文討論了..."
Include specific technical terms, data, solution names, or viewpoints
Preserve key numbers and metrics (like performance improvement percentages, user counts, version numbers, etc.)
If article involves comparison/selection, point out comparison objects and conclusion
Goal: Reader spends 30 seconds reading summary and can decide whether worth spending 10 minutes on original

Output format (internal):

Article summaries = [
  {
    index: 0,
    titleTranslated: "翻譯後的標題" or "Original Title",
    summary: "Summary content...",
    reason: "Recommendation reason..."
  },
  ...
]

Step 7: Generate Today's Highlights

Based on the top 10 articles, write a 3-5 sentence "Today's Highlights" summary.

Requirements:

Extract 2-3 main trends or topics in today's tech world
Don't list article by article, do macro summarization
Style: concise and powerful, like news lead
Language: Use the language specified by user (zh-tw or en)

Output: Plain text string (not JSON, not markdown formatting).

Step 8: Build Markdown Report

Generate a complete Markdown report with the following structure:

Report Structure

# 📰 AI 部落格每日精選 — YYYY-MM-DD
(or "# 📰 AI Blog Daily Digest — YYYY-MM-DD" if language is en)

> 來自 Karpathy 推薦的 ${metadata.totalFeeds} 個頂級技術部落格，AI 精選 Top ${articles.length}
(or "> From ${metadata.totalFeeds} top tech blogs recommended by Karpathy, AI-selected Top ${articles.length}" if language is en)

## 📝 今日看點
(or "## 📝 Today's Highlights" if language is en)

${highlights}

---

## 🏆 今日必讀
(or "## 🏆 Today's Must-Read" if language is en)

${top 3 articles with medals 🥇🥈🥉, each showing:
  - Translated title (bold)
  - English title (link) — source · relative time · category emoji + label
  - Summary (blockquote)
  - Recommendation reason
  - Keywords
}

---

## 📊 數據概覽
(or "## 📊 Data Overview" if language is en)

| 掃描來源 | 抓取文章 | 時間範圍 | 精選 |
(or "| Scanned Sources | Fetched Articles | Time Range | Selected |" if language is en)
|:---:|:---:|:---:|:---:|
| ${successFeeds}/${totalFeeds} | ${totalArticles} → ${filteredArticles} | ${hours}h | **${selected} 篇** |
(or "**${selected} articles**" if language is en)

### 分類分佈
(or "### Category Distribution" if language is en)

${Mermaid pie chart showing category distribution}

### 高頻關鍵詞
(or "### Frequent Keywords" if language is en)

${Mermaid bar chart showing top 12 keywords}

<details>
<summary>📈 純文字關鍵詞圖（終端機友善）</summary>
(or "<summary>📈 Plain Text Keyword Chart (Terminal-Friendly)</summary>" if language is en)

${ASCII bar chart with █ and ░ characters}

</details>

### 🏷️ 話題標籤
(or "### 🏷️ Topic Tags" if language is en)

${Tag cloud: top 20 keywords, top 3 bold with counts}

---

## ${Category emoji} ${Category label}

${For each category group, list all articles:
  ### N. ${titleTranslated}
  
  [${title}](${link}) — **${sourceName}** · ${relative time} · ⭐ ${totalScore}/30
  
  > ${summary}
  
  🏷️ ${keywords}
  
  ---
}

${Repeat for all categories sorted by article count descending}

Helper Functions You Need

Relative Time:

For zh-tw:
- <60 min: "X 分鐘前"
- <24 hours: "X 小時前"
- <7 days: "X 天前"
- ≥7 days: ISO date (YYYY-MM-DD)
For en:
- <60 min: "X minutes ago"
- <24 hours: "X hours ago"
- <7 days: "X days ago"
- ≥7 days: ISO date (YYYY-MM-DD)

Mermaid Pie Chart (category distribution):

pie showData
    title "文章分類分佈" (or "Article Category Distribution" if en)
    "🤖 AI / ML" : ${count}
    "🔒 安全 / Security" : ${count}
    ...

Mermaid Bar Chart (top 12 keywords):

xychart-beta horizontal
    title "高頻關鍵詞" (or "Frequent Keywords" if en)
    x-axis ["keyword1", "keyword2", ...]
    y-axis "出現次數" 0 --> ${maxCount + 2} (or "Occurrences" if en)
    bar [count1, count2, ...]

ASCII Bar Chart (top 10 keywords, terminal-friendly):

keyword1  │ ████████████████████ 15
keyword2  │ ████████████░░░░░░░░ 12
...

Tag Cloud (top 20 keywords):

**keyword1**(15) · **keyword2**(12) · **keyword3**(10) · keyword4(8) · keyword5(7) · ...

(First 3 bold)

Category Metadata:

For zh-tw:

ai-ml       → 🤖 AI / ML
security    → 🔒 安全
engineering → ⚙️ 工程
tools       → 🛠 工具 / 開源
opinion     → 💡 觀點 / 雜談
other       → 📝 其他

For en:

ai-ml       → 🤖 AI / ML
security    → 🔒 Security
engineering → ⚙️ Engineering
tools       → 🛠 Tools / Open Source
opinion     → 💡 Opinions / Miscellaneous
other       → 📝 Other

Step 9: Write Report to File

Save the Markdown report:

mkdir -p ./output
OUTPUT_FILE="./output/digest-$(date +%Y%m%d).md"

# Use Write tool to create the file
Write($OUTPUT_FILE, <markdown content>)

Step 10: Save Configuration

Save the current parameters to config file:

mkdir -p ~/.tomtom-daily-digest
cat > ~/.tomtom-daily-digest/config.json << 'EOF'
{
  "timeRange": <hours>,
  "topN": <topN>,
  "language": "<zh-tw|en>",
  "lastUsed": "<ISO timestamp>"
}
EOF

Step 11: Present Results to User

Show a summary: