360 Secure SkillHub
360 安全技能中心

Search Synthesis Expert

通过顺序推理将用户问题分解为可独立检索的子问题并制定搜索策略,借助 Playwright 模拟浏览完成多源信息采集,最后经去重、交叉验证与可信度评估综合为带来源标注的报告,适用于深度研究、竞争分析与事实核查。

openlark
openlark
热度 16.02
最近更新 2026-06-02 10:43:25
分类:搜索依赖:联网浏览搜索 / 代码执行 / 内部调用大模型应用场景:研究 / 摘要总结 / 报告撰写 / 市场分析

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 search-synthesis-expert
Metadata
name
search-synthesis-expert
description
Uses sequential-thinking to decompose tasks and formulate search strategies, browser automation (Playwright) for multi-source information search and collection, and final review of synthesized results. Suitable for deep research, competitive analysis, technical investigation, and fact-checking requiring multi-source synthesis.

Search Synthesis Expert

Expert at searching and synthesizing information. Executes tasks in a three-phase process: Plan → Collect → Review.

Workflow

User query → Phase 1: Decompose & Plan → Phase 2: Browser search & collect → Phase 3: Review & Synthesize → Output report

Phase 1: Decompose & Plan (sequential-thinking)

Use step-by-step reasoning to decompose tasks and formulate search strategies:

  1. Clarify goal: Understand the user's core question and information boundaries
  2. Decompose sub-problems: Break complex questions into independently searchable sub-questions
  3. Formulate search strategy: Determine keywords, target sites, and search order for each sub-problem
  4. Prioritize: Order searches by information importance

Example output:

Task decomposition:
1. Sub-question A → Search keywords: {keywords} → Target: {site}
2. Sub-question B → Search keywords: {keywords} → Target: {site}
3. Sub-question C → Search keywords: {keywords} → Target: {site}

Phase 2: Browser Search & Collect (Playwright)

Use Playwright to simulate user browsing for search and information collection:

  1. Open search engine or target site
  2. Enter search keywords, simulating real user behavior
  3. Browse search results, click relevant links
  4. Extract page content, collect key information
  5. Multi-source comparison: Get information from different sources for the same question
  6. Record sources: Save each piece of information's URL for traceability

Core operations:

# Example operation flow (pseudocode)
page.goto("https://www.google.com")
page.fill("input[name='q']", keyword)
page.press("input[name='q']", "Enter")
page.wait_for_selector("div#search")
results = page.query_selector_all("h3")
# Click relevant results → Extract content → Record sources

Phase 3: Review & Synthesize (sequential-thinking)

Use step-by-step reasoning to review collected information:

  1. Deduplicate: Merge duplicate information from multiple sources
  2. Cross-validate: Compare consistency across sources, flag contradictions
  3. Credibility assessment: Evaluate information credibility based on source authority
  4. Structured synthesis: Organize final output by logical relationships
  5. Gap marking: Flag information gaps that could not be filled

Output Format

═══════════════════════════════════
Search & Synthesis Report

📋 Query
{original user question}

🔍 Search Strategy
{sub-problems and keyword list}

📚 Sources
| # | Source | Summary | Credibility |
|---|--------|---------|-------------|
| 1 | {URL} | {summary} | ⭐⭐⭐ |
| 2 | {URL} | {summary} | ⭐⭐ |

📝 Synthesized Conclusion
{structured complete answer}

⚠️ Notes
- {information limitations}
- {uncovered aspects}
- {suggestions for further research}
═══════════════════════════════════

Best Practices

Principle Description
Multi-source verification Each key fact verified from at least 2 independent sources
Source traceability Each piece of information tagged with source URL for backtracking
Timeliness first Prioritize latest information, note information date
Breadth first Search same question from different angles (official docs, communities, blogs, news)
Honest labeling Clearly mark unfound information as "not found," never fabricate

Notes

  • Set reasonable User-Agent and delays when using Playwright to avoid anti-bot measures
  • For sites requiring login, mark as "login required"
  • Non-deterministic information (predictions, opinions) must note the source's position
  • If search cannot cover all sub-problems, mark missing parts in the report