research（研究）

Research Skill

CRITICAL: Use the Research CLI for Multi-Source Research

For ANY research request, use the research CLI via Bash:

# All sources in parallel (RECOMMENDED - fastest)
~/.claude/bin/research --all "your research query"

# Individual sources
~/.claude/bin/research --perplexity "query"    # Fast web search
~/.claude/bin/research --gemini "query"        # Multi-perspective
~/.claude/bin/research --claude "query"        # Detailed analysis (slower)

DO NOT use WebSearch directly for research requests - use the CLI above.

Three Research Modes

QUICK RESEARCH MODE

• Trigger: User says "quick research" or simple queries
• Method: Single research --all "query" call
• Time: ~15-30 seconds
• Result: 3 source results (Perplexity, Claude, Gemini)

STANDARD RESEARCH MODE (Default)

• Trigger: Default for most research requests
• Method: 3 parallel Bash calls with different query angles
• Time: ~30-45 seconds
• Result: 9 source results (3 angles × 3 sources)

EXTENSIVE RESEARCH MODE

• Trigger: User says "extensive research" or "deep dive"
• Method: 8 parallel Bash calls covering diverse angles
• Time: ~45-90 seconds
• Result: 24 source results (8 angles × 3 sources)

Standard Research Workflow

Step 1: Decompose the Question

Break the user's question into 3 focused sub-questions:

• Core facts and recent developments
• Context, implications, and analysis
• Contrarian views, controversies, or edge cases

Step 2: Launch Parallel Research (ONE MESSAGE)

CRITICAL: Use a SINGLE message with multiple Bash tool calls for parallel execution

# Launch all 3 in parallel (single message with 3 Bash calls)
~/.claude/bin/research --all "sub-question 1 focusing on core facts"
~/.claude/bin/research --all "sub-question 2 focusing on implications"
~/.claude/bin/research --all "sub-question 3 focusing on controversies"

Step 3: Parse JSON Results

Each call returns JSON:

{
  "source": "perplexity|claude|gemini",
  "query": "the query",
  "timestamp": "ISO timestamp",
  "success": true,
  "content": "research findings...",
  "citations": ["url1", "url2"],
  "duration_ms": 1234
}

For --all, returns array of 3 results.

Step 4: Synthesize Findings

Confidence Levels:

• HIGH CONFIDENCE: Corroborated by 2+ sources
• MEDIUM CONFIDENCE: Found by one source, seems reliable
• LOW CONFIDENCE: Single source, needs verification

Structure:

## Key Findings

### [Topic Area 1]
**High Confidence:**
- Finding X (Sources: perplexity, claude)

**Medium Confidence:**
- Finding Z (Source: gemini only)

## Source Attribution
- **Perplexity**: [web/current events]
- **Claude**: [detailed analysis]
- **Gemini**: [multi-perspective synthesis]

## Conflicting Information
- [Note any disagreements]

Extensive Research Workflow

Step 1: Generate 8 Diverse Angles

• Core facts and current state
• Historical context and evolution
• Technical deep-dive
• Practical implications and applications
• Contrarian views and criticisms
• Future predictions and trends
• Cross-domain connections
• Edge cases and unusual perspectives

Step 2: Launch 8 Parallel Queries

# All 8 in parallel (single message)
~/.claude/bin/research --all "angle 1: core facts about [topic]"
~/.claude/bin/research --all "angle 2: historical context of [topic]"
~/.claude/bin/research --all "angle 3: technical deep-dive on [topic]"
~/.claude/bin/research --all "angle 4: practical implications of [topic]"
~/.claude/bin/research --all "angle 5: criticisms around [topic]"
~/.claude/bin/research --all "angle 6: future predictions for [topic]"
~/.claude/bin/research --all "angle 7: cross-domain connections to [topic]"
~/.claude/bin/research --all "angle 8: edge cases of [topic]"

API Keys Required

Located in ~/.claude/.env:

Critical Rules

Parallel Execution

• ✅ Launch ALL research queries in ONE message (parallel Bash calls)
• ✅ Each query covers a different angle
• ❌ DON'T launch sequentially (kills speed benefit)
• ❌ DON'T wait between queries

Error Handling

• If a source fails, proceed with successful results
• Note failures in the final report
• Check success: false in JSON for errors

Research Metrics Template

Include at end of every research report:

**📈 RESEARCH METRICS:**
- **Mode:** [Quick/Standard/Extensive]
- **Total Queries:** [X] (angles × 3 sources)
- **Sources:** Perplexity, Claude, Gemini
- **Confidence Level:** [High/Medium/Low] ([%])

Example: Standard Research

User: "Research quantum computing developments"

Kai executes:

# 3 parallel Bash calls in ONE message
~/.claude/bin/research --all "quantum computing recent breakthroughs 2026"
~/.claude/bin/research --all "quantum computing practical applications timeline"
~/.claude/bin/research --all "quantum computing challenges limitations"

Result: 9 source queries (3 angles × 3 sources) in ~30-45 seconds.