multi-search（多重搜索）

Multi-Topic Deep Research Skill

Overview

Intelligent multi-topic deep research tool that automatically analyzes materials and generates systematic research documents. Supports arbitrary material input, launching multiple independent research agents in parallel for deep retrieval, forming a concise research knowledge base.

Core Principles:

• Only perform information retrieval, summarization, and professional expression transformation
• No adding facts, no fabricating information
• Focus on single problems, concise output, just enough to solve the problem
• Universal design, applicable to legal, business, technical, academic, and other fields

Trigger Conditions

Trigger with /multi-search command, or when users request:

• Deep research on multiple related topics
• Systematic information retrieval
• Multi-perspective analysis integration
• Structured research report generation

Input Format

Method 1: File-based

/multi-search @document-path.md

Method 2: Direct Paste

/multi-search
[Paste material content]

Method 3: Specify Topics

/multi-search
Project: [Project Name]
Research Topics:
1. [Topic One]
2. [Topic Two]
3. [Topic Three]

Processing Flow

Phase 1: Analysis Preparation

1. Read input materials
2. Extract research topic list

- With clear topics: Use directly - Without clear topics: Auto-extract from materials

1. Topic splitting principles:

- Clear direction: Each topic corresponds to a unique retrieval direction - Avoid overlap: Ensure no duplicate retrieval keywords between topics - Focus on problems: Each topic solves one specific problem

1. Determine project name and output location

Phase 2: Output Directory Detection

Detect project structure by priority:

1. Priority detection: output/ directory -> Use output/[project-name]/
2. Secondary detection: Current working directory -> Use ./[project-name]/
3. Fallback: User's current directory -> Use ./research/

Create directory: [output-dir]/03 - Deep Research/

Phase 3: Parallel Deep Research

Launch an independent general-purpose research agent for each research topic.

Context Transfer (Main Agent -> Research Agent):

• Project key information (background, objectives, core problems)
• Complete topic list and retrieval scope for each topic
• Assigned keyword directions (basis for avoiding duplicates)
• Specific requirement background

Deduplication Mechanism:

Each research agent must follow this process before starting retrieval:

1. Pre-retrieval Declaration:

- Declare in current context: "I will search [Keyword A, Keyword B] for researching [Topic Name]" - Wait for main agent confirmation of no duplicates before starting

1. Main Agent Review:

- Check if the agent's declared keywords duplicate assigned directions - If duplicates found, notify the agent to pivot to other directions

1. Dynamic Adjustment:

- If a direction is already covered by other agents, pivot to related but different angles - Record adjusted retrieval directions

Deep Retrieval Requirements:

• 4-6 rounds of deep retrieval
• Auto-select WebSearch (discovery) or WebFetch (get full content)
• Differentiated keywords, ensuring each agent covers unique angles

Document Generation:

• Focus on solving a single core problem
• Concise and clear, just enough to solve the problem
• Include key source links
• Directly usable conclusions and recommendations

Phase 4: Integration Output

1. Generate research overview document (000.Research-Overview.md)
2. Integrate core findings from all research agents
3. Create inter-document navigation links
4. Add comprehensive recommendations and immediate action list

Output Format

Directory Structure

[output-dir]/
└── [project-name]/
    └── 03 - Deep Research/
        ├── 000.Research-Overview.md
        ├── YYMMDD [Research Topic One].md
        ├── YYMMDD [Research Topic Two].md
        └── ...

Overview Document Format

# [Project Name] Deep Research Overview

**Generated**: YYYY-MM-DD
**Research Method**: N independent research agents, each conducting 4-6 rounds of deep retrieval
**Total Retrieval Rounds**: XX+ rounds
**Total Document Size**: XX KB

---

## Research Deliverables

### N Concise Research Reports Completed

| No. | Research Topic | File Size | Core Value |
|-----|----------------|-----------|------------|
| 01 | [Topic One](./YYMMDD%20Topic-One.md) | XX KB | Brief description |

---

## Core Findings

### Finding 1: [Most Important Finding]

**Basis**: [Brief explanation]

**Conclusion**: [Specific conclusion]

---

## Comprehensive Recommendations

### I. Strategic Recommendations

**Recommended Approach**: [Specific approach]

### II. Immediate Action List

- [ ] Action item 1
- [ ] Action item 2

Detailed Research Document Format

# [Research Topic Title]

**Generated**: YYYY-MM-DD
**Research Depth**: XX+ rounds of deep retrieval, covering XXXX, XXXX, XXXX

---

## Core Conclusions

[Most important findings and conclusions, 2-3 paragraphs, thorough and detailed]

---

## I. [Main Content One]

### (1) Subsection

Body paragraph. Use inline link format for source citations:
- According to [Source Name](https://link)...
- Based on [Material](https://link)...

---

## II. [Main Content Two]

[Continue structured content]

---

## III. Application Recommendations

### (1) Key Recommendations

**Content**: [Specific content]

### (2) Precautions

⚠️ [Warning point]

Link Specifications

Core Principle

All source links must be embedded inline at corresponding positions in the text

✅ Correct:
According to [research report](https://link)...

❌ Incorrect:
According to some report...
(References listed separately at end)

Link Notation Conventions

• 🔗 -> General web resources
• 📚 -> Academic literature
• 🏛️ -> Institutional websites
• 📄 -> Data sources

Document Naming Conventions

Numbering System

• 00. - Research overview
• 01-09. - Core research
• 10-19. - Important research
• 20+. - Extended research

Title Guidelines

• ✅ Use concise titles
• ✅ Avoid special characters
• ✅ Length within 15 words
• ✅ Clearly reflect research subject

Quality Standards

Research Agent Quality

• Focus on single problem: Each research agent solves only one core problem
• Retrieval depth: 4-6 rounds of retrieval (just enough)
• Concise output: Clear and concise, just enough to solve the problem
• Key citations: Cite key sources (just enough)
• Directly usable: Provide directly actionable conclusions and recommendations

Document Quality Standards

• Clear structure: Chapter titles with clear hierarchy
• Coherent narrative: Paragraph-style narrative, avoid excessive listing
• Accurate links: All links embedded inline at corresponding positions
• Consistent format: Follow unified format specifications
• Strong actionability: Provide specific steps, tools, commands

Precautions

Prohibited Actions

• ❌ Do not create sub-subdirectories (e.g., "reference-materials/")
• ❌ Do not generate separate executive summary files
• ❌ Do not use excessive bullet-point listing format
• ❌ Do not list references separately at document end
• ❌ Do not add redundant progress tracking sections

Recommended Practices

• ✅ Use narrative paragraph expression
• ✅ Embed links at corresponding text positions
• ✅ Keep research overview concise
• ✅ Provide specific action recommendations
• ✅ Mark clear document numbers

Dependencies

This skill relies on Claude Code built-in tools, no additional configuration needed:

• WebSearch: Search discovery
• WebFetch: Get full content
• Task: Launch independent research agents

Changelog