Conduct multiple rounds of searching, cross-validation, and perspective aggregation analysis on complex topics.
Core Capabilities
- Multi-round follow-up questioning: discover new questions from initial results and continue digging deeper
- Multi-source cross-validation: important information is confirmed by at least 2 independent sources
- Perspective aggregation analysis: identify positions, analyze interest relationships, and track changes over time
- Structured output: integrate fragmented information into a knowledge system
Configuration Parameters
Users can specify via natural language:
| Parameter | Value | Default | Description |
|---|---|---|---|
| Depth | 1-5 | 3 | Controls the number of follow-up rounds, 5 = most in-depth |
| Breadth | 1-5 | 3 | Controls source coverage, 5 = most extensive |
| Output | full/summary/table | full | Report format |
| Language | cn/en/both | both | Information source language preference |
Example: "Help me conduct deep research on AI Agent, depth 5, output a complete report"
Execution Flow
Phase 1: Intent Parsing and Planning
- Parse the user query and identify the core question
- Determine the query type:
- Fact verification: need to verify the truth of a claim
- Perspective comparison: need to understand the views of different positions
- Technical research: need to deeply understand a certain technology/product
- Market analysis: need to understand the market landscape and competitor situation
- Generate a search plan based on the type and parameters
Phase 2: Broad Search and Discovery
Use WebSearch for parallel searching, covering Chinese and English sources:
Search Strategy：
- Generate 2-3 queries from different angles for each core question
- Positive query + negative query + neutral query
- Search separately in Chinese and English
Chinese platforms(use site: to restrict):
- Zhihu zhihu.com - in-depth discussion
- Weibo weibo.com - real-time public opinion
- WeChat Official Accounts mp.weixin.qq.com - in-depth articles
- Bilibili bilibili.com - views of younger groups
English platforms：
- Reddit reddit.com - community discussion
- Twitter/X x.com - real-time updates
- Hacker News news.ycombinator.com - technical discussion
- GitHub github.com - technical documentation
Identify follow-up points：
- Vague source ("according to reports") → ask for the specific source
- Data without time → ask for the time
- Contradictory sources → ask for the reason for the difference
- One-sided statement → ask for the views of other parties
- Technical claim without evidence → ask for cases
Phase 3: Deep Verification and Aggregation
- Cross-verify important claims (≥2 independent sources)
- Use WebFetch to retrieve the full text of key pages (only high-value pages)
- Identify the stance and interests of opinion holders
- Determine whether to continue follow-up questioning (return to Phase 2)
Opinion Analysis Framework：
Stakeholder classification:
- Producers: technology developers, manufacturers, service providers
- Consumers: end users, purchase decision-makers
- Regulators: government agencies, industry associations
- Commentators: media, analysts, academic researchers
- Competitors: direct competitors, alternative solution providers
Stance identification signals:
- Strong support: absolute wording, emphasis only on advantages, statements by interested parties
- Strong opposition: threatening wording, emphasis only on risks, statements by threatened parties
- Neutral: balanced phrasing, provision of data and qualifications, independent institutions
Credibility assessment:
- Academic papers/official documents > authoritative media > industry analysis > company statements > social media
Loop Control：
- Depth 1: 1 round of search, quick overview
- Depth 3: up to 3 rounds, standard research
- Depth 5: up to 5 rounds, in-depth investigation
- Termination condition: new information is less than 10% or all follow-up points have been addressed
Phase 4: Structured Output
Select a template according to the output format parameter; see reference.md。
Complete Report Structure：
# [主题] 深度调研报告
## 执行摘要
> 3-5 句话概括核心发现
## 研究概述
- 研究问题 / 范围 / 方法 / 置信度
## 核心发现
### 发现 1
- 结论 + 证据支撑 + 存疑点
## 观点图谱（如适用）
| 立场 | 代表方 | 核心论点 | 利益分析 | 证据强度 |
## 信息缺口与不确定性
## 来源清单
Executive summary: output only the summary section, ≤500 words
Comparison table: suitable for product/solution comparison, presented in table form
Tool Usage
| Tool | Purpose | Precautions |
|---|---|---|
| WebSearch | Search for information | 5-10 times per round, with different keyword combinations |
| WebFetch | Retrieve full text | Use only for high-value pages to save tokens |
| Task | Parallel subtasks | Use when there are multiple independent subtopics |
Constraints
- Fully automatic execution, no mid-process confirmation from the user
- After each round of search, output brief progress (X leads found, Y questions followed up)
- Clearly mark information that cannot be verified
- All citations must include source links
- Mark the time of the information and pay attention to timeliness
Detailed Reference
Follow-up decision rules, complete stakeholder classification, detailed version of the report template, platform search strategy: see reference.md