Deep Research

Deep Research Skill

Core Principles

Every research task MUST follow:

1. Plan first, execute later - Don't start collecting data immediately
2. Multi-dimensional exploration - Go deep, don't stay shallow
3. Verify data - Sources must be cited, important data cross-verified
4. Dig into anomalies - Unusual points are often the most valuable
5. Maintain explorer's mindset - Curiosity-driven, always ask "why"

Auto-Trigger Conditions

MUST use this skill when user message involves:

• Any "research X", "analyze X", "investigate X" requests
• Questions about a topic/industry/company/technology/person
• Data-driven decision making
• Comparison and evaluation tasks
• Deep understanding of a field
• Trends, outlook, development direction topics

Research Execution Flow

Phase 1: Research Planning (REQUIRED)

Before any actual research, MUST complete these steps:

1.1 Clarify Research Objectives

Ask yourself:

• What does the user really want to know? (beyond the surface question)
• What is the ultimate value of this research?
• What form should the output be? (report/data/recommendations)

1.2 Brainstorm Research Dimensions (MUST list at least 5)

Think systematically about these aspects:

Important: The 6th "Unique Angles" dimension is key to creating differentiated value - don't skip it.

1.3 Determine Priorities

Categorize:

• Must Have: Core questions that must be answered
• Should Have: Value-adding supplementary analysis
• Nice to Have: Unique insights and extended thinking

1.4 Define Quality Criteria

Auto-generate review_criteria for subsequent review, including:

• Which dimensions must be covered for completeness
• What data quality standards to meet
• What depth of analysis is expected
• What sections the report must contain

Phase 2: Data Collection & Verification

Universal Verification Rules

1. Prioritize reliable data sources

   • Professional tools (e.g., akshare-stocks, akshare-a-shares)
   • Official websites and documentation
   • Authoritative media and institutions

2. Cross-verify important data

   • At least 2 sources for confirmation
   • If conflicts exist, analyze reasons

3. MUST cite data sources and timestamps

   • Source: Where the data came from
   • Time: Data timeliness
   • If historical data, clearly mark it

Domain-Specific Rules

Phase 3: Deep Exploration (Core Value Phase)

For each research dimension, execute this flow:

┌─────────────────────────────────────┐
│  1. Basic Information Collection    │
│     - Gather fundamental data/facts │
└─────────────┬───────────────────────┘
              ▼
┌─────────────────────────────────────┐
│  2. Anomaly Identification (KEY)    │
│     - What data looks unusual?      │
│     - What trends deserve attention?│
│     - What's commonly overlooked?   │
└─────────────┬───────────────────────┘
              ▼
┌─────────────────────────────────────┐
│  3. Deep Investigation (anomalies)  │
│     - Why is this happening?        │
│     - What's the underlying cause?  │
│     - What impacts/chain effects?   │
└─────────────┬───────────────────────┘
              ▼
┌─────────────────────────────────────┐
│  4. Form Insights                   │
│     - What does this mean?          │
│     - What value for the user?      │
│     - What actionable suggestions?  │
└─────────────────────────────────────┘

Phase 4: Comprehensive Report

Report Structure (MUST include)

1. Executive Summary

   • 2-3 sentences summarizing core conclusions
   • Let readers quickly grasp key points

2. Key Findings

   • Each finding must have data support
   • Cite data sources and timestamps
   • Distinguish facts from inferences

3. Deep Analysis

   • In-depth exploration of anomalies
   • Analysis from different angles
   • Unique insights and observations

4. Data Appendix

   • Sources of all cited data
   • Retrieval timestamps
   • Data limitations explained

5. Risk Alerts/Limitations

   • Data limitations
   • Analysis assumptions
   • Possible biases

6. Conclusions & Recommendations

   • Clear recommendations based on research
   • Directions for further exploration
   • Points requiring ongoing attention

Review Guidelines (For Main Agent)

When reviewing research results, check these dimensions:

Coverage Check

• Does it answer the user's core question?
• Does it cover all planned dimensions?
• Are there obvious missing important aspects?

Depth Check

• Does analysis stay at surface-level data?
• Are anomalies/interesting points explored deeply?
• Are there unique insights and observations?
• Does it ask "why"?

Data Quality Check

• Are data sources cited?
• Is important data cross-verified?
• Are timestamps clear?
• Is historical data clearly marked?

Logic Check

• Are conclusions supported by data?
• Is the reasoning sound?
• Does it distinguish facts from inferences?

When Rejecting, MUST Provide

If deciding to reject, must give:

1. Specific Issue: Which dimension is insufficient/inaccurate
2. Missing Content: What else should be analyzed
3. Exploration Directions: 2-3 specific improvement directions
4. Improvement Guidance: Tell Sub Agent exactly how to improve

Example rejection feedback:

REJECT

Issue: Analysis stays at surface level, only lists data without analyzing reasons.

Missing dimensions:
- No comparison with industry averages
- No analysis of historical trend changes
- No exploration of reasons for anomalous data points

Improvement directions:
1. Compare with data from other companies in the same industry
2. Analyze trends over the past 3 years
3. Deeply explore why XXX metric is abnormally high

Specific suggestion: XXX data is significantly higher than industry average,
this is worth exploring from market positioning, cost structure,
and competitive advantage perspectives.

Quality Criteria Template

When using delegate_and_review, reference this quality criteria template:

Research report must satisfy:

1. Structural Completeness
   - Contains summary, findings, analysis, data appendix, conclusions
   - Each section has substantial content, not empty

2. Data Quality
   - All data has cited sources
   - Key data verified from at least 2 sources
   - Data timeliness is clear

3. Analysis Depth
   - Not just listing data, must have analysis
   - Anomalies are explored in depth
   - Has unique insights, not generic statements

4. Practical Value
   - Conclusions are clear and actionable
   - Recommendations are specific and implementable
   - Risk alerts are explicit