| name | web-content-fetcher |
| description | Extract article content from any URL as clean Markdown. Uses Scrapling script as primary method (with auto fast→stealth fallback), Jina Reader as alternative for simple pages. Preserves headings, links, images, lists, and code blocks. Use this skill whenever the user wants to fetch, read, extract, scrape, or summarize content from a URL — including blog posts, news articles, WeChat articles (微信公众号), documentation pages, or any web page. Also trigger when the user says things like "帮我读一下这篇文章", "抓取这个网页", "提取正文", or "read this page for me".
|
Web Content Fetcher
Given a URL, return its main content as clean Markdown — headings, links, images, lists, code blocks all preserved.
Note: This skill extracts content + remote image URLs. If the user wants an "offline" copy (download images to local disk and rewrite links), add a post-processing step (not included by default in this skill).
Extraction Strategy
Always try one method per URL — don't cascade blindly. Pick the right one upfront.
URL
│
├─ 1. Scrapling script (preferred)
│ Run fetch.py — check the domain routing table to decide fast vs --stealth.
│ Works for most sites. Returns clean Markdown directly.
│
└─ 2. Jina Reader (fallback — only if Scrapling fails or dependencies not installed)
web_fetch("https://r.jina.ai/<url>")
Free tier: 200 req/day. Fast (~1-2s), good Markdown output.
Does NOT work for: WeChat (403), some Chinese platforms.
Scrapling script
python3 <SKILL_DIR>/scripts/fetch.py "<url>" [max_chars] [--stealth]
<SKILL_DIR> is the directory where this SKILL.md lives. Resolve it before calling the script.
The script has two modes built in:
- Default (fast): HTTP fetch, ~1-3s, works for most sites
--stealth: Headless browser, ~5-15s, for JS-rendered or anti-scraping sites
When run without --stealth, the script automatically falls back to stealth if the fast result has too little content. So you rarely need to specify --stealth manually — the only reason to force it is when you already know the site needs it (see routing table), which saves the initial fast attempt.
Domain Routing
Use this table to pick the right mode on the first call:
| Domain | Command | Why |
|---|
mp.weixin.qq.com | fetch.py <url> --stealth | JS-rendered content |
zhuanlan.zhihu.com | fetch.py <url> --stealth | Anti-scraping + JS |
juejin.cn | fetch.py <url> --stealth | JS-rendered SPA |
sspai.com | fetch.py <url> | Static HTML |
blog.csdn.net | fetch.py <url> | Static HTML |
ruanyifeng.com | fetch.py <url> | Static blog |
openai.com | fetch.py <url> | Static HTML |
blog.google | fetch.py <url> | Static HTML |
| Everything else | fetch.py <url> | Auto-fallback handles it |
Script Options
python3 <SKILL_DIR>/scripts/fetch.py "https://sspai.com/post/73145"
python3 <SKILL_DIR>/scripts/fetch.py "https://mp.weixin.qq.com/s/xxx" --stealth
python3 <SKILL_DIR>/scripts/fetch.py "https://example.com/article" 15000
python3 <SKILL_DIR>/scripts/fetch.py "https://example.com" --json
Install Dependencies
First use only — the script checks and tells you if anything is missing:
pip install scrapling html2text
If on system-managed Python (macOS/Linux), add --break-system-packages or use a venv.
Failure Rules
- Same URL fails once → give up, tell the user "unable to extract content from this URL"
- Do not retry — each failed call wastes context tokens
WeChat-specific gotchas
- WeChat often uses lazy-loaded images where the real URL is in
data-src and src is a tiny placeholder.
- The extractor script normalizes these to real
src URLs before running html2text.
- If you ever see Markdown image lines that contain weird URL-encoded SVG fragments (e.g.
...%3Csvg...%3E) appended after the closing ) of an image, it means the placeholder leaked into Markdown parsing; update fix_lazy_images() in scripts/fetch.py to remove/replace placeholder data:image/svg+xml src values.
