npx skills add https://github.com/cyranob/web-forager --skill web-research此技能引导你完成结构化的深度调研工作流程:多角度搜索、选择性全文抓取,并综合成一份精炼且附有引用的报告。
搜索 — 使用 ddgs 库运行 Python:
from ddgs import DDGS
results = DDGS().text(query="your query", max_results=8)
for r in results:
print(r["title"], r["href"], r["body"])
如果 ddgs 未安装:pip install ddgs
抓取 — 直接调用 Jina Reader 以获取 URL 的纯净 Markdown 内容:
curl -s "https://r.jina.ai/https://example.com"
或在 Python 中:
import requests
content = requests.get(f"https://r.jina.ai/{url}").text
无需 API 密钥。
如果会话中提供了 MCP 搜索/抓取工具(例如 、 或类似工具),请优先使用这些工具而非上述方法——它们速度更快且已集成。
广告位招租
在这里展示您的产品或服务
触达数万 AI 开发者,精准高效
mcp__web_forager__searchmcp__duckduckgo__search在开始搜索之前,花点时间理解实际被询问的内容:
不要只搜索一次。使用不同的查询表述方式运行 2-3 次搜索,以获得更广泛的覆盖范围。思考:
运行所有搜索,收集结果。先不要抓取 URL——先评估摘要片段。
从所有搜索结果中,根据以下标准挑选 3-5 个最有希望的 URL:
在抓取之前,向用户简要解释你的 URL 选择——这可以建立信任,并让他们在你选择不当时能够纠正你。
使用上述 Jina Reader 方法抓取每个选定的 URL。使用 Markdown 格式(默认)。如果抓取失败,则跳过该 URL 并在你的来源部分注明。
对于非常长的页面,专注于最相关的部分,而不是包含所有内容。
使用以下格式撰写一份结构化的报告。目标是生成一份用户真正可以使用的文档——不是意识流,也不是引用列表。
根据问题调整报告的规模。并非每个问题都需要一份包含 5 个部分的报告。
当用户只需要一个事实、日期、比较或简短答案时:
## [答案标题]
用 2-4 句话直接回答,关键事实放在最前面。
**来源:** [标题](url), [标题](url)
# [调研标题]
## 摘要
用 2-4 句话直接回答核心问题的 TL;DR(太长不看版总结)。
## 关键发现
- 你学到的最重要、最具体的事实的要点列表
- 每个要点应独立成句——避免"根据来源 X,……"
- 在可能的情况下包含数字、日期、具体细节
## 来源
1. [标题](url) — 一行描述此来源的贡献
# [调研标题]
## 摘要
用 2-4 句话直接回答核心问题的 TL;DR(太长不看版总结)。
## 关键发现
- 包含数字、日期、具体细节的具体要点
## [主题部分 — 2-4 个部分,以其涵盖内容命名]
针对每个主要方面进行更深入阐述的段落。
在行内引用来源,格式为 [来源名称](url)。
## 局限性与空白
你未能找到的信息、不确定之处、用户应进一步挖掘的地方。
## 来源
1. [标题](url) — 一行描述此来源的贡献
2. [标题](url) — ...
运用你的判断。目标是:为用户提供适量的信息,而不是最大量的信息。
每周安装数
1
代码仓库
GitHub 星标数
4
首次出现
1 天前
安全审计
安装于
amp1
cline1
opencode1
cursor1
kimi-cli1
codex1
This skill guides you through a structured deep research workflow: multi-angle search, selective full-page fetching, and synthesis into a polished, cited report.
Search — run Python with the ddgs library:
from ddgs import DDGS
results = DDGS().text(query="your query", max_results=8)
for r in results:
print(r["title"], r["href"], r["body"])
If ddgs isn't installed: pip install ddgs
Fetch — call Jina Reader directly to get a URL as clean markdown:
curl -s "https://r.jina.ai/https://example.com"
Or in Python:
import requests
content = requests.get(f"https://r.jina.ai/{url}").text
No API key needed.
If MCP search/fetch tools are available in the session (e.g., mcp__web_forager__search, mcp__duckduckgo__search, or similar), prefer those over the above — they're faster and already wired in.
Before searching, take a moment to understand what's actually being asked:
Don't search just once. Run 2–3 searches with different query framings to get broader coverage. Think about:
Run all searches, collect results. Don't fetch URLs yet — evaluate snippets first.
From all the search results, pick the 3–5 most promising URLs based on:
Briefly explain your URL selection to the user before fetching — this builds trust and lets them redirect you if you've chosen poorly.
Fetch each selected URL using the Jina Reader method above. Use markdown format (default). If a fetch fails, skip that URL and note it in your sources section.
For very long pages, focus on the most relevant sections rather than including everything.
Write a structured report using the format below. The goal is a document the user can actually use — not a stream of consciousness, not a list of quotes.
Scale the report to match the question. Not every question needs a 5-section report.
When the user just needs a fact, a date, a comparison, or a short answer:
## [Answer title]
Direct answer in 2–4 sentences, with the key fact front and center.
**Sources:** [Title](url), [Title](url)
# [Research title]
## Summary
2–4 sentence TL;DR that answers the core question directly.
## Key findings
- Bullet points of the most important, concrete things you learned
- Each bullet should stand alone — avoid "according to source X, ..."
- Include numbers, dates, specifics where available
## Sources
1. [Title](url) — one line describing what this source contributed
# [Research title]
## Summary
2–4 sentence TL;DR that answers the core question directly.
## Key findings
- Concrete bullet points with numbers, dates, specifics
## [Topic sections — 2–4 sections, named for what they cover]
Prose paragraphs going deeper on each major aspect.
Cite sources inline as [Source Name](url).
## Limitations & gaps
What you couldn't find, what's uncertain, where the user should dig further.
## Sources
1. [Title](url) — one line describing what this source contributed
2. [Title](url) — ...
Use your judgment. The goal is: give the user the right amount of information, not the maximum amount.
Weekly Installs
1
Repository
GitHub Stars
4
First Seen
1 day ago
Security Audits
Gen Agent Trust HubPassSocketFailSnykWarn
Installed on
amp1
cline1
opencode1
cursor1
kimi-cli1
codex1
通过 LiteLLM 代理让 Claude Code 对接 GitHub Copilot 运行 | 高级变通方案指南
31,600 周安装