一个 OpenClaw Agent 技能，让你的 AI 助手变成一个有方法论的研究员。不是简单地丢搜索结果，而是根据问题自动调整搜索深度——从快速事实核查到多源深度调研——最终输出一份信源分级的结构化报告。
这是一个搜索方法论技能，不是开箱即用的搜索引擎。它教你的 OpenClaw agent 怎么搜才搜得好：
- 自适应深度 — 简单问题快速收敛，复杂问题自动多轮挖掘
- 信源分级 — 每个来源标注 🟢🟡🔴（官方 → 媒体 → 社区）
- 交叉验证 — 关键结论至少两个独立来源互相印证
- 静默执行 — 用户只看到一条最终报告，中间过程不刷屏
search/
├── SKILL.md # 核心技能定义（"大脑"）
├── scripts/
│ └── cn-social.py # 中文社媒搜索（302.AI）
└── references/
├── source-priority.md # 信源可靠性分级体系
└── community-gates.md # 社区来源使用规范
⚠️ 这个技能不能直接用。 它依赖未包含在内的外部工具。本仓库作为参考实现分享——展示搜索技能的结构设计和有效的方法论。
| 依赖 | 作用 | 是否包含？ |
|---|---|---|
| OpenClaw | Agent 运行时 | ❌（需单独安装） |
web_fetch |
网页全文提取 | ✅（OpenClaw 内置） |
| Exa search skill | 主要网页搜索引擎 | ❌ |
| bird CLI | X/Twitter 社区搜索 | ❌ |
| 302.AI API key | 中文社媒（知乎、B站、小红书、微信公众号、TikTok） | 部分（脚本包含，需自备 API key） |
技能遵循四阶段流水线：
- 意图分析（静默） — 将查询分类（事实核查/信息收集/动态追踪/舆情探查），规划信源策略
- 探查 — 1-2 次轻量搜索建立认知框架。如果答案已经明确，直接跳到报告
- 证据收集循环 — 迭代：识别最大信息缺口 → 针对性搜索 → 更新证据图谱 → 重复。当核心问题有可靠来源支撑，或连续两轮无新信息时停止
- 信源评级与报告 — 所有来源标注可靠性等级，输出结构化报告
- 替换搜索后端 — 把
SKILL.md中的 Exa 引用换成你的搜索工具（Tavily、SearXNG、Perplexity、Google Custom Search 等） - 替换或移除社区搜索 — 把
bird引用换成你有的社区搜索工具，或直接删掉 - 保留方法论 — 信源分级体系、交叉验证规则和自适应深度逻辑与具体工具无关
An OpenClaw agent skill that turns your AI assistant into a methodical researcher. Instead of dumping raw search results, it automatically calibrates search depth based on the question — from a quick fact-check to a multi-source investigation — and delivers a single, source-graded report.
This is a search methodology skill, not a turnkey search engine. It teaches your OpenClaw agent how to search well:
- Adaptive depth — simple questions get quick answers; complex ones trigger multi-round evidence gathering
- Source grading — every source is rated 🟢🟡🔴 (official → media → community)
- Cross-verification — key claims require at least two independent sources
- Silent execution — the user sees one final report, not a stream of intermediate searches
search/
├── SKILL.md # Core skill definition (the "brain")
├── scripts/
│ └── cn-social.py # Chinese social media search (302.AI)
└── references/
├── source-priority.md # Source reliability ranking system
└── community-gates.md # Rules for using community sources
⚠️ This skill will NOT work out of the box. It relies on external tools that are not included. This repo is shared as a reference implementation — to show how a search skill can be structured and what methodology works well.
| Dependency | Role | Included? |
|---|---|---|
| OpenClaw | Agent runtime | ❌ (install separately) |
web_fetch |
Full-text page extraction | ✅ (built into OpenClaw) |
| Exa search skill | Primary web search engine | ❌ |
| bird CLI | X/Twitter community search | ❌ |
| 302.AI API key | Chinese social media (Zhihu, Bilibili, Xiaohongshu, WeChat, TikTok) | Partial (script included, API key required) |
The skill follows a 4-stage pipeline:
- Intent Analysis (silent) — classifies the query (fact-check / info gathering / trend tracking / sentiment probe) and plans the source strategy
- Recon — 1–2 lightweight searches to establish the knowledge landscape. If the answer is already clear, skip to the report
- Evidence Collection Loop — iterative: identify the biggest information gap → targeted search → update evidence map → repeat. Stops when core questions are answered with reliable sources, or when two consecutive rounds yield no new information
- Source Grading & Report — all sources tagged with reliability levels, structured report output
- Replace the search backend — swap Exa references in
SKILL.mdwith your preferred search tool (Tavily, SearXNG, Perplexity, Google Custom Search, etc.) - Replace or remove the social search — swap
birdreferences with whatever community search you have, or remove those sections - Keep the methodology — the source grading system, cross-verification rules, and adaptive depth logic are tool-agnostic
MIT