research-external（研究外部）

External Research Workflow

Research external sources (documentation, web, APIs) for libraries, best practices, and general topics.

Note: The current year is 2025. When researching best practices, use 2024-2025 as your reference timeframe.

Invocation

/research-external <focus> [options]

Question Flow (No Arguments)

If the user types just /research-external with no or partial arguments, guide them through this question flow. Use AskUserQuestion for each phase.

Phase 1: Research Type

question: "What kind of information do you need?"
header: "Type"
options:
  - label: "How to use a library/package"
    description: "API docs, examples, patterns"
  - label: "Best practices for a task"
    description: "Recommended approaches, comparisons"
  - label: "General topic research"
    description: "Comprehensive multi-source search"
  - label: "Compare options/alternatives"
    description: "Which tool/library/approach is best"

Mapping:

• "How to use library" → library focus
• "Best practices" → best-practices focus
• "General topic" → general focus
• "Compare options" → best-practices with comparison framing

Phase 2: Specific Topic

question: "What specifically do you want to research?"
header: "Topic"
options: []  # Free text input

Examples of good answers:

• "How to use Prisma ORM with TypeScript"
• "Best practices for error handling in Python"
• "React vs Vue vs Svelte for dashboards"

Phase 3: Library Details (if library focus)

If user selected library focus:

question: "Which package registry?"
header: "Registry"
options:
  - label: "npm (JavaScript/TypeScript)"
    description: "Node.js packages"
  - label: "PyPI (Python)"
    description: "Python packages"
  - label: "crates.io (Rust)"
    description: "Rust crates"
  - label: "Go modules"
    description: "Go packages"

Then ask for specific library name if not already provided.

Phase 4: Depth

question: "How thorough should the research be?"
header: "Depth"
options:
  - label: "Quick answer"
    description: "Just the essentials"
  - label: "Thorough research"
    description: "Multiple sources, examples, edge cases"

Mapping:

• "Quick answer" → --depth shallow
• "Thorough" → --depth thorough

Phase 5: Output

question: "What should I produce?"
header: "Output"
options:
  - label: "Summary in chat"
    description: "Tell me what you found"
  - label: "Research document"
    description: "Write to thoughts/shared/research/"
  - label: "Handoff for implementation"
    description: "Prepare context for coding"

Mapping:

• "Research document" → --output doc
• "Handoff" → --output handoff

Summary Before Execution

Based on your answers, I'll research:

**Focus:** library
**Topic:** "Prisma ORM connection pooling"
**Library:** prisma (npm)
**Depth:** thorough
**Output:** doc

Proceed? [Yes / Adjust settings]

Focus Modes (First Argument)

Options

Workflow

Step 1: Parse Arguments

Extract from user input:

FOCUS=$1           # library | best-practices | general
TOPIC="..."        # from --topic
DEPTH="shallow"    # from --depth (default: shallow)
OUTPUT="doc"       # from --output (default: doc)
LIBRARY="..."      # from --library (optional)
REGISTRY="npm"     # from --registry (default: npm)

Step 2: Execute Research by Focus

Focus: library

Primary tool: nia-docs - Find API documentation, usage patterns, code examples.

# Semantic search in package
(cd $CLAUDE_OPC_DIR && uv run python -m runtime.harness scripts/mcp/nia_docs.py \
  --package "$LIBRARY" \
  --registry "$REGISTRY" \
  --query "$TOPIC" \
  --limit 10)

# If thorough depth, also grep for specific patterns
(cd $CLAUDE_OPC_DIR && uv run python -m runtime.harness scripts/mcp/nia_docs.py \
  --package "$LIBRARY" \
  --grep "$TOPIC")

# Supplement with official docs if URL known
(cd $CLAUDE_OPC_DIR && uv run python -m runtime.harness scripts/mcp/firecrawl_scrape.py \
  --url "https://docs.example.com/api/$TOPIC" \
  --format markdown)

Thorough depth additions:

• Multiple semantic queries with variations
• Grep for specific function/class names
• Scrape official documentation pages

Focus: best-practices

Primary tool: perplexity-search - Find recommended approaches, patterns, anti-patterns.

# AI-synthesized research (sonar-pro)
(cd $CLAUDE_OPC_DIR && uv run python scripts/mcp/perplexity_search.py \
  --research "$TOPIC best practices 2024 2025")

# If comparing alternatives
(cd $CLAUDE_OPC_DIR && uv run python scripts/mcp/perplexity_search.py \
  --reason "$TOPIC vs alternatives - which to choose?")

Thorough depth additions:

# Chain-of-thought for complex decisions
(cd $CLAUDE_OPC_DIR && uv run python scripts/mcp/perplexity_search.py \
  --reason "$TOPIC tradeoffs and considerations 2025")

# Deep comprehensive research
(cd $CLAUDE_OPC_DIR && uv run python scripts/mcp/perplexity_search.py \
  --deep "$TOPIC comprehensive guide 2025")

# Recent developments
(cd $CLAUDE_OPC_DIR && uv run python scripts/mcp/perplexity_search.py \
  --search "$TOPIC latest developments" \
  --recency month --max-results 5)

Focus: general

Use ALL available MCP tools - comprehensive multi-source research.

Step 2a: Library documentation (nia-docs)

(cd $CLAUDE_OPC_DIR && uv run python -m runtime.harness scripts/mcp/nia_docs.py \
  --search "$TOPIC")

Step 2b: Web research (perplexity)

(cd $CLAUDE_OPC_DIR && uv run python scripts/mcp/perplexity_search.py \
  --research "$TOPIC")

Step 2c: Specific documentation (firecrawl)

# Scrape relevant documentation pages found in perplexity results
(cd $CLAUDE_OPC_DIR && uv run python -m runtime.harness scripts/mcp/firecrawl_scrape.py \
  --url "$FOUND_DOC_URL" \
  --format markdown)

Thorough depth additions:

• Run all three tools with expanded queries
• Cross-reference findings between sources
• Follow links from initial results for deeper context

Step 3: Synthesize Findings

Combine results from all sources:

1. Key Concepts - Core ideas and terminology
2. Code Examples - Working examples from documentation
3. Best Practices - Recommended approaches
4. Pitfalls - Common mistakes to avoid
5. Alternatives - Other options considered
6. Sources - URLs for all citations

Step 4: Write Output

Output: doc (default)

Write to: thoughts/shared/research/YYYY-MM-DD-{topic-slug}.md

---
date: {ISO timestamp}
type: external-research
topic: "{topic}"
focus: {focus}
sources: [nia, perplexity, firecrawl]
status: complete
---

# Research: {Topic}

## Summary
{2-3 sentence summary of findings}

## Key Findings

### Library Documentation
{From nia-docs - API references, usage patterns}

### Best Practices (2024-2025)
{From perplexity - recommended approaches}

### Code Examples

// Working examples found

## Recommendations

- {Recommendation 1}
- {Recommendation 2}

## Pitfalls to Avoid

- {Pitfall 1}
- {Pitfall 2}

## Alternatives Considered

| Option | Pros | Cons |
| --- | --- | --- |
| {Option 1} | ... | ... |

## Sources

- [{Source 1}](https://github.com/parcadei/continuous-claude-v3/blob/HEAD/.claude/skills/research-external/%7Burl1%7D)
- [{Source 2}](https://github.com/parcadei/continuous-claude-v3/blob/HEAD/.claude/skills/research-external/%7Burl2%7D)

Output: handoff

Write to: thoughts/shared/handoffs/{session}/research-{topic-slug}.yaml

---
type: research-handoff
ts: {ISO timestamp}
topic: "{topic}"
focus: {focus}
status: complete
---

goal: Research {topic} for implementation planning
sources_used: [nia, perplexity, firecrawl]

findings:
  key_concepts:
    - {concept1}
    - {concept2}

  code_examples:
    - pattern: "{pattern name}"
      code: |
        // example code

  best_practices:
    - {practice1}
    - {practice2}

  pitfalls:
    - {pitfall1}

recommendations:
  - {rec1}
  - {rec2}

sources:
  - title: "{Source 1}"
    url: "{url1}"
    type: {documentation|article|reference}

for_plan_agent: |
  Based on research, the recommended approach is:
  1. {Step 1}
  2. {Step 2}
  Key libraries: {lib1}, {lib2}
  Avoid: {pitfall1}

Step 5: Return Summary

Research Complete

Topic: {topic}
Focus: {focus}
Output: {path to file}

Key findings:
- {Finding 1}
- {Finding 2}
- {Finding 3}

Sources: {N} sources cited

{If handoff output:}
Ready for plan-agent to continue.

Error Handling

If an MCP tool fails (API key missing, rate limited, etc.):

1. Log the failure in output: tool_status: nia: success perplexity: failed (rate limited) firecrawl: skipped
2. Continue with other sources - partial results are valuable
3. Set status appropriately:

- complete - All requested tools succeeded - partial - Some tools failed, findings still useful - failed - No useful results obtained

1. Note gaps in findings: ## Gaps - Perplexity unavailable - best practices section limited to nia results

Examples

Library Research (Shallow)

/research-external library --topic "dependency injection" --library fastapi --registry py_pi

Best Practices (Thorough)

/research-external best-practices --topic "error handling in Python async" --depth thorough

General Research for Handoff

/research-external general --topic "OAuth2 PKCE flow implementation" --depth thorough --output handoff

Quick Library Lookup

/research-external library --topic "useEffect cleanup" --library react

Integration with Other Skills

Required Environment

• NIA_API_KEY or nia server in mcp_config.json
• PERPLEXITY_API_KEY in environment or ~/.claude/.env
• FIRECRAWL_API_KEY and firecrawl server in mcp_config.json

Notes

• NOT for codebase exploration - Use research-codebase or scout for that
• Always cite sources - Include URLs for all findings
• 2024-2025 timeframe - Focus on current best practices
• Graceful degradation - Partial results better than no results