ai-daily-report

name	ai-daily-report
description	生成每日 AI/技术情报日报。从 Hacker News 和 HuggingFace Papers 抓取高质量内容，按主题分类、评级，生成中文日报。默认保存为本地 Markdown， 可选发布到飞书知识库。 当用户说"生成日报"、"AI日报"、"今日技术资讯"、"每日情报"、"/ai-daily"、 "tech daily"、"今天有什么新闻"、"抓一下今天的内容"时触发此技能。 即使用户只是随口问"今天 HN 上有什么有意思的"，也应该触发。

AI Daily Report

从 Hacker News 和 HuggingFace Papers 抓取当日高质量技术内容，按分类整理为中文日报。默认保存为本地 Markdown 文件，可选发布到飞书等平台。

Workflow Overview

1. Check cache  -->  2. Crawl sources  -->  3. Deduplicate & filter
      |                    |                        |
      v                    v                        v
4. Score & categorize  -->  5. Generate Markdown  -->  6. Save locally
                                                       │
                                                       └─→  7. (Optional) Publish to Feishu Wiki

Step 1: Check Cache

Before crawling, check what's already been collected today to avoid duplicate work.

python3 ~/.claude/skills/ai-daily-report/scripts/cache.py get --date $(date +%Y-%m-%d)

If entries exist for today, use them as a starting point. Only re-crawl sources that haven't been fetched yet.

Step 2: Crawl Sources

Read data/resource.json (sibling to this SKILL.md) for source configuration. Each source has a crawl_difficulty that determines which tool to use:

Easy sources (curl) -- Hacker News

HN has a clean Firebase API. Fetch top and best stories, then individual items:

# Get top 30 story IDs
curl -s 'https://hacker-news.firebaseio.com/v0/topstories.json' | python3 -c "
import json, sys
ids = json.load(sys.stdin)[:30]
print(json.dumps(ids))
"

# Fetch a single item
curl -s "https://hacker-news.firebaseio.com/v0/item/{id}.json"

Each item returns: {id, title, url, score, by, time, descendants, type}.

Filter: keep items with score >= 50.

For items that link to GitHub issues, blog posts, or other pages, use WebFetch to get a summary of the linked content. This is what turns a bare HN link into a useful report item.

Medium sources (WebFetch) -- HuggingFace Papers

Use WebFetch on https://huggingface.co/papers with a prompt to extract:
- Paper title
- Authors / institution
- Upvote count
- One-line summary
- Paper URL

Filter: keep papers with upvotes >= 10.

Hard sources (Playwright) -- reserved for future sources

If a source requires browser automation (JS rendering, login, infinite scroll), use the Playwright MCP tools to navigate, wait for render, and extract content.

Step 3: Deduplicate & Cache

After crawling, write all new items to cache:

python3 ~/.claude/skills/ai-daily-report/scripts/cache.py put '{
  "url": "https://...",
  "title": "...",
  "source": "hackernews",
  "score": 150,
  "date": "2026-04-13",
  "summary": "..."
}'

The cache script automatically deduplicates by URL and title.

Step 4: Score & Categorize

Separate the items into two independent tracks before categorizing:

Track A -- Hacker News items: Assign each to a topic category (see references/format-guide.md for the taxonomy). These become the main body of the report.

Track B -- HuggingFace Papers: Do NOT assign these to topic categories. They all go into one dedicated section: AI 研究论文（HuggingFace Papers 今日精选）. Keep them as a separate list throughout the pipeline.

For each item in both tracks, determine:

1. Rating -- pick the appropriate rating label (重要程度/有趣程度/实用程度/关注程度) and level (极高/高/中高/中)
2. Summary -- write a Chinese summary (2-5 sentences for HN items, 1-2 sentences for HF papers)

Rating guidelines:

• 极高: Paradigm shifts, critical security issues, >500 HN points, top HuggingFace paper
• 高: Significant tools/releases, notable research, 200-500 HN points
• 中高: Useful tools, interesting findings, 100-200 HN points
• 中: Worth knowing, 50-100 HN points

For HN items that link to substantial content (blog posts, papers, GitHub issues), use WebFetch to read the linked page and write a more informed summary. A bare title is not enough -- the value of the report is in the synthesis.

Step 5: Generate Markdown

Read references/format-guide.md for the exact output format. Generate the full Markdown report following that template.

Key rules:

• Title: 每日技术情报日报 · {date} · 共 {N} 条高质量精选内容
• 15-25 items total (HN + HF combined). If you have more than 25, raise the HN score threshold or drop the lowest-rated items until you're within range. Prefer quality over quantity.
• Group HN items by topic category with --- separators (these form the main body)
• After all HN topic sections, place HuggingFace papers in one dedicated section: ## AI 研究论文（HuggingFace Papers 今日精选）. Papers must NOT appear inside the topic categories above -- they are always in this final section before the footer.
• Footer with source attribution and generation date

Save the Markdown to a local path. Ask the user where to save, or use a sensible default like ./ai-daily-{date}.md or the user's preferred output directory.

Step 6: Save Locally

The report is always saved as a local Markdown file first. Ask the user for their preferred save path if not specified. Common patterns:

# Default: current directory or user-specified path
./ai-daily-{date}.md

# Or a structured directory
{output_dir}/ai-report/ai-daily-{date}.md

Step 7: Publish (Optional)

This step is optional. Only execute if the user explicitly requests publishing to a platform.

Option A: Feishu (Lark) Wiki

If the user wants to publish to Feishu, they need to provide:

1. Wiki Space ID — from their Feishu knowledge base URL
2. Root Node Token — the parent node under which reports are organized
3. feishu-cli installed and authenticated (feishu-cli auth login)

Reports are organized by month under the root node:

知识库首页 (root_node_token)
  └── {YYYY}年{M}月 (month node)
       ├── 每日技术情报日报 · 2026-04-07
       ├── 每日技术情报日报 · 2026-04-08
       └── ...

Publishing Steps

7a. Find or create the monthly node

feishu-cli wiki node list {space_id} --parent-node-token {root_node_token}

Look for a node titled like 2026年4月. If it doesn't exist, create it:

feishu-cli wiki node create {space_id} \
  --obj_type docx \
  --parent_node_token {root_node_token} \
  --title "{YYYY}年{M}月"

7b. Create the daily report document

feishu-cli wiki node create {space_id} \
  --obj_type docx \
  --parent_node_token {monthly_node_token} \
  --title "每日技术情报日报 · {date} · 共 {N} 条高质量精选内容"

7c. Import the Markdown content

feishu-cli doc import /path/to/ai-daily-{date}.md \
  --document-id {obj_token} \
  --title "每日技术情报日报 · {date} · 共 {N} 条高质量精选内容"

Or write content directly:

feishu-cli doc add {obj_token} /path/to/ai-daily-{date}.md --content-type markdown

7d. Return the URL

https://{tenant}.feishu.cn/wiki/{node_token}

Option B: Other Platforms

The skill generates standard Markdown. Users can integrate with any platform that accepts Markdown:

• Notion (via API or import)
• Confluence (via Markdown import)
• GitHub Wiki / Pages
• Any CMS with Markdown support

The skill only handles Feishu natively via feishu-cli. For other platforms, provide the local Markdown path for the user to handle.

Crawling Strategy: Use Sub-Agents for Parallelism

For best performance, dispatch source crawling in parallel:

1. Agent 1: Crawl Hacker News (top + best stories via curl, then WebFetch for interesting links)
2. Agent 2: Crawl HuggingFace Papers (WebFetch)

Each agent should return a JSON array of items with: url, title, source, score, date, category_hint, summary.

Then merge results, deduplicate, categorize, generate the report, and publish.

Error Handling

• If feishu-cli fails with auth errors, suggest the user run the feishu-cli-auth skill
• If a source is unreachable, skip it and note in the report footer
• If cache.py fails, continue without caching (the report still works)
• If the monthly Wiki node creation fails with permission errors, fall back to creating the doc under the root node

Command Reference

Action	Command
Check cache	python3 ~/.claude/skills/ai-daily-report/scripts/cache.py get --date YYYY-MM-DD
Add to cache	python3 ~/.claude/skills/ai-daily-report/scripts/cache.py put '<json>'
List wiki nodes	feishu-cli wiki node list {space_id}
Create wiki node	feishu-cli wiki node create {space_id} --obj_type docx --parent_node_token <token> --title "..."
Get node info	feishu-cli wiki node get <node_token>
Import markdown	feishu-cli doc import /tmp/file.md --title "..."
Write to doc	feishu-cli doc add <doc_id> /tmp/file.md --content-type markdown