360 Secure SkillHub
360 安全技能中心

Free Web Browser

该 Skill 指导通过 OpenClaw 的 web_search 与 web_fetch 完成联网检索:先把请求改写为具体查询,先检索再抓取 1 至 3 个来源页面,交叉核对关键事实,最后输出带来源链接与时效说明的摘要。搜索假定由 Ollama 提供,无需额外搜索 API 密钥;若运行时未暴露 web_search,则无法联网检索。

prantikmedhi
prantikmedhi
热度 50.56
最近更新 2026-05-11 17:44:17
分类:搜索依赖:第三方 API应用场景:研究 / 摘要总结 / 推荐 / 笔记整理

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add free-web-browser
Metadata
name
free-web-browser
description
Browse web through OpenClaw web_search and web_fetch with Ollama as search provider, no external API key required. Use when user asks to search web, verify current facts, compare sources, open result pages, summarize articles, or browse live web content cheaply through local or Ollama-backed search instead of paid search APIs or full browser automation.

Free Web Browser

Use web_search for discovery, then web_fetch for page content.

Quick workflow

  1. Rewrite vague request into concrete search query.
  2. Run web_search first.
  3. Pick 1 to 3 promising URLs.
  4. Run web_fetch on source pages you actually need.
  5. Cross-check important claims across more than one source.
  6. Summarize findings with links, caveats, and date sensitivity when relevant.

Query patterns

Prefer short, specific queries.

  • Current fact: site:docs.openclaw.ai web_search ollama
  • Product compare: ollama web search brave searxng comparison
  • Official docs first: site:github.com OR site:docs.openclaw.ai <topic>
  • Recent news: <topic> 2026

If first query weak, tighten scope instead of fetching random pages.

Tool choice

web_search

Use for:

  • finding sources
  • checking whether topic has changed recently
  • locating official docs, repos, blog posts, issues
  • getting top candidate URLs before deeper reading

web_fetch

Use for:

  • extracting readable page content
  • summarizing article, doc, or blog post
  • verifying exact wording from source page
  • pulling details after search already found target URL

Prefer extractMode: "markdown" unless plain text is better for logs or machine-like parsing.

Quality rules

  • Prefer official docs, project repos, vendor docs, standards bodies, and primary sources.
  • Do not trust snippets alone for important claims. Fetch source page.
  • For time-sensitive topics, say info may change and mention recency.
  • When sources disagree, say so plainly.
  • Avoid unnecessary fetch spam. Search once, fetch only best pages.

Ollama-specific guidance

This skill assumes web_search is configured to use Ollama-backed search, so no separate search API key is needed.

  • Use normal web_search tool. Do not shell out to Ollama manually unless debugging config.
  • If search fails, suspect provider setup before blaming query.
  • If Ollama search returns thin snippets, fetch source pages for substance.
  • If current runtime does not expose web_search, this skill cannot browse live web.

Failure handling

  • No results: broaden query once, then try official-site search.
  • Low-quality results: add site filters or product/version terms.
  • Fetch blocked or empty: choose alternate source from search results.
  • Ambiguous request: ask one clarifying question only after one good-faith search attempt.

Output pattern

Default answer shape:

  • direct answer first
  • 2 to 5 bullets with evidence
  • source links
  • brief uncertainty note if needed

Read next

For troubleshooting and example prompts, read references/ollama-web-browser-notes.md.