Claude Code 的七阶段深度研究系统。多源验证、思维图谱方法论、医疗保健/金融/法律/市场调研领域叠加。
概览
Claude Code 的七阶段研究系统，产出决策级、可审计、抗幻觉的研究成果。 1. 下载此仓库为ZIP 2.解压为 ~/.claude/skills/deep-research/ 3.重启Claude代码 示例：- “深入研究量子计算现状” - “深度研究SaaS定价最佳实践” - “深入研究欧洲的人工智能监管” 系统会自动对你的问题进行分类： 所有研究资料都保存在 /RESEARCH/[topic_name]/： 1. - 将无来源声明标记为[需要来源] 2.- 关键声明需要独立核实 3.- 5条引用文章 1 报告 = 1 来源 4.- 记录并解释，不要隐藏 5.
概述（AI 翻译）
Claude Code 的七阶段研究系统，产出决策级、可审计、抗幻觉的研究成果。
这有什么不同？
| 传统人工智能研究 | 深度研究 |
|---|---|
| 单次搜索查询 | 5-10次并行搜索 |
| 信任第一个结果 | 用2+来源验证 |
| 无引用 | 所有引用的声明均附有网址 |
| 隐藏的矛盾 | 冲突的记录与解释 |
| 通用答案 | 假设驱动调查 |
安装
选项1：克隆到技能目录
选项2：手动安装
- 下载此仓库的ZIP
- 提取到
~/.claude/skills/deep-research/ - 重启 Claude 代码
用途
只需说：“”深入研究[你的主题]“
示例：
- “深入研究量子计算现状”
- “SaaS定价的深度研究最佳实践”
- “彻底研究欧洲的人工智能监管”
研究类型
系统会自动对你的问题进行分类：
| 类型 | 时间 | 用例 |
|---|---|---|
| A：查找1-2分钟 | 权威来源的单一事实 | |
| B：合成15-30分钟 | 汇总多个来源 |
本页由 AI 翻译，技术细节请以原文为准。
README
以下为原文 README 节选，完整内容请查看原文。
Deep Research Skill
A 7-phase research system for Claude Code that produces decision-grade, auditable, hallucination-resistant research outputs.
What Makes This Different?
| Traditional AI Research | Deep Research |
|---|---|
| Single search query | 5-10 parallel searches |
| Trust first result | Verify with 2+ sources |
| No citations | Every claim cited with URL |
| Hidden contradictions | Conflicts documented and explained |
| Generic answers | Hypothesis-driven investigation |
Installation
Option 1: Clone to Skills Directory
cd ~/.claude/skills
git clone https://github.com/YOUR_USERNAME/deep-research-skill.git deep-research
Option 2: Manual Installation
- Download this repository as a ZIP
- Extract to
~/.claude/skills/deep-research/ - Restart Claude Code
Usage
Simply say: “Deep research [your topic]”
Examples:
- “Deep research the current state of quantum computing”
- “Deep research best practices for SaaS pricing”
- “Research AI regulation in Europe thoroughly”
Research Types
The system automatically classifies your question:
| Type | Time | Use Case |
|---|---|---|
| A: Lookup | 1-2 min | Single fact from authoritative source |
| B: Synthesis | 15-30 min | Aggregating multiple sources |
| C: Analysis | 30-60 min | Judgment and multiple perspectives |
| D: Investigation | 2-4 hours | Novel questions, conflicting evidence |
The 7 Phases
0. Classify → What type of research is this?
1. Scope → What exactly are we researching?
1.5 Hypothesize → What are the likely answers?
2. Plan → How will we find information?
3. Query → Execute parallel searches
4. Triangulate → Verify claims across sources
5. Synthesize → Combine into coherent findings
6. QA → Verify citations and claims
7. Package → Deliver structured output
Output Structure
All research is saved to /RESEARCH/[topic_name]/:
RESEARCH/
└── [topic_name]/
├── README.md # Navigation guide
├── 00_research_contract.md # Scope agreement
├── 01_research_plan.md # Subquestions, queries
├── 02_query_log.csv # All searches
├── 03_source_catalog.csv # Sources with ratings
├── 04_evidence_ledger.csv # Claims → Sources
├── 08_report/
│ ├── 00_executive_summary.md
│ ├── 01_findings.md
│ └── 02_recommendations.md
└── 09_qa/
└── qa_report.md
Key Principles
- No claim without evidence - Mark unsourced claims as
[Source needed] - 2-source rule - Critical claims need independent verification
- Independence matters - 5 articles citing 1 report = 1 source
- Contradictions are data - Document and explain, don’t hide
- Track everything - Query logs, source catalogs, evidence ledgers
File Structure
deep-research/
├── SKILL.md # Main skill file推荐工具
换一个关键词，或者移除筛选条件。
安装
npx skillfish add standardhuman/deep-research-skill