deep-research - SKILL.md Agent Skill

name: deep-research description: Conduct comprehensive multi-source research with AI-powered analysis. Use when investigating complex topics, gathering information from multiple sources, or synthesizing research findings. model: opus allowed-tools: Bash, Task, Read, Write, Edit, MultiEdit, WebSearch, WebFetch, Grep, Glob, TodoWrite argument-hint: [optional-focus-area]

Deep Research Command

Conduct systematic, multi-source research on specified topics using AI-powered analysis and cross-referenced insights.

🎯 Purpose & Scope

What this command does NOT do:

Generate research without source attribution
Create content without verification

When to REJECT:

Topic is better suited for simple web search

🔄 Workflow

ultrathink: you'd perform the following steps

Step 1: Planning

Analyze Requirements
- Parse research topic from $ARGUMENTS
- Identify optional focus areas
- Determine research scope and boundaries
- Select Research Strategy based on topic type:
  - Technical/Scientific: Prioritize academic papers, documentation, code repos
  - Historical: Focus on primary sources, archives, temporal progression
  - Current Events: Emphasize recent sources, news, real-time updates
  - Theoretical/Conceptual: Balance foundational texts with recent developments
  - Product/Market: Industry reports, user feedback, competitive analysis
- Plan source discovery strategy aligned with selected approach
Initialize Research Infrastructure
- Create research workspace: research_$ARGUMENTS_$(date +%Y%m%d)
- Initialize source tracker: source_tracker.md
- Set up finding templates for cross-source insights
- Check for existing research sessions: Look for source_tracker.md in workspace
  - If exists: Resume from last state (check status fields)
  - Auto-save progress to tracker after each source analysis
  - Tracker serves as both documentation and checkpoint
  - Resume capability: Continue from any 'pending' or 'analyzing' sources
Identify Applicable Workflows & Standards
- Identify any relevant workflows for research processes
- Identify any related documentation standards
- Note: MUST follow any matching workflows and standards
Delegation Decision
- Plan parallel source analysis tasks
- Identify specialized research agents needed
- Define synthesis coordination points
Risk Assessment
- Identify potential source reliability issues
- Plan verification strategies
- Note information gaps

Step 2: Execution

Workflow Compliance
- MUST follow workflows identified in Phase 1
- Apply research methodology standards
- Reference specific workflow files when applicable
Primary Implementation
- Dispatch subagents for iterative source analysis
- Each agent analyzes assigned sources for:
  - Key information relevant to topic
  - Cross-source patterns and themes
  - Contradictions and debates
  - NEW INSIGHTS requiring further research
- Update source tracker with analysis status
- CRITICAL: When finding new insights, discover additional sources
Cross-Source Finding Format
- Pattern/Theme identification
- Evidence compilation from multiple sources
- Implication analysis for research topic
- Source attribution with locations
Iterative Deep Research Strategy
- Discovery Triggers: When subagents identify:
  - Unfamiliar concepts or terminology
  - References to related work
  - Gaps in understanding
  - Contradictory information requiring more context
- Action Protocol:
  1. Add discovered sources to tracker immediately
  2. Spawn new research tasks for high-relevance discoveries
  3. Update finding files with connection chains
  4. Document research depth (initial → level 2 → level 3)
- Depth Limits: Maximum 3 levels of iterative discovery
- Priority Management: High-relevance discoveries analyzed first
Edge Case Handling
- No sources found: Expand search parameters
- Duplicate sources: Check content similarity before analysis
- Conflicting information: Document all perspectives
- Subagent timeout: Mark as failed, continue with available data
- Discovery overflow: Prioritize by relevance + credibility
- Circular references: Document but don't re-analyze
- Early termination: If high-confidence convergent evidence achieved

Step 3: Verification

Quality Assurance
- Verify all sources marked as analyzed have summaries
- Confirm discovered sources are tracked and processed
- Cross-reference findings across multiple sources
- Identify contradictions between sources
- Flag gaps in research coverage
- Validate discovery chains (source → insight → new source)
Consolidation Tasks
- Merge similar findings from different subagents
- Resolve conflicting information with source priority
- Apply Synthesis Framework:
  - Convergent Evidence: Multiple sources supporting same conclusion
  - Divergent Perspectives: Document contrasting viewpoints with context
  - Confidence Scoring: Weight findings by source credibility + convergence
  - Knowledge Gaps: Explicitly mark where evidence is insufficient
- Create synthesis notes for complex topics with confidence levels
Source Validation
- Confirm source reliability and credibility
- Check for citation accuracy
- Verify quotes and references
Side Effect Validation
- Clean temporary sketchpad files
- Ensure all findings traceable to sources
- Verify workspace organization

Step 4: Reporting

Output Format:

[✅/❌] Research: $ARGUMENTS

## Summary
- Initial sources: [count]
- **Discovered sources: [count] (across [N] levels)**
- Total sources analyzed: [count]
- **Duplicates detected: [count]**
- Key findings: [count]
- Research gaps: [count]
- **Overall confidence**: [High/Medium/Low]
- **Discovery depth reached: [1-3 levels]**
- **Credibility distribution**: Academic:[N] | Official:[N] | Industry:[N] | Community:[N]

## Executive Summary
[3-5 sentence overview of key discoveries]

## Key Findings
### Finding 1: [Title] (Confidence: High/Medium/Low)
- **Evidence**: [Sources and quotes with credibility indicators]
- **Convergence**: [N sources agree] | [M sources disagree]
- **Significance**: [Why this matters]

### Finding 2: [Title] (Confidence: High/Medium/Low)
[Continue pattern...]

## Thematic Analysis
### Theme A: [Emergent Pattern]
[Synthesis across multiple findings]

## Debates and Contradictions
[Areas where sources disagree]

## Discovery Chains
### Chain 1: [Initial Topic → Discovery]
- Started with: [initial source]
- Led to discovery of: [concept/source]
- Further revealed: [deeper insight]

### Chain 2: [Another Discovery Path]
[Continue pattern...]

## Research Gaps
[What wasn't found or needs further investigation]

## References
[Complete source list with access information, grouped by discovery level]
### Level 1 (Initial Sources)
- [Source 1]
- [Source 2]

### Level 2 (Discovered from initial research)
- [Source A] ← discovered from [Source 1]
- [Source B] ← discovered from [Source 2]

### Level 3 (Deep discoveries)
- [Source X] ← discovered from [Source A]

📝 Examples

Basic Research

/deep-research "quantum computing applications"
# Conducts comprehensive research on quantum computing uses

Focused Research

/deep-research "machine learning" "healthcare applications"
# Researches ML specifically in healthcare context

Complex Multi-Domain Research

/deep-research "climate change impact on global supply chains"
# Analyzes intersection of climate and supply chain topics

Iterative Discovery Research

/deep-research "emerging AI safety frameworks"
# Initial search finds 3 sources
# Agent A discovers reference to "constitutional AI" → adds 2 new sources
# Agent B finds mention of "RLHF techniques" → triggers deep dive with 4 sources
# Agent C identifies "red teaming methodologies" → spawns specialized search
# Result: 15 total sources analyzed across 3 discovery levels

Parallel Source Analysis

/deep-research "blockchain scalability solutions"
# Automatically delegates to:
#   - Agent A: Academic papers analysis
#   - Agent B: Industry reports (parallel)
#   - Agent C: Technical documentation (parallel)
#   - Agent D: Synthesis and consolidation

Error Case Handling

/deep-research "classified-information"
# Error: Topic requires restricted access
# Suggestion: Refine to publicly available information
# Alternative: Focus on general concepts instead