Searching Deeply

Core Strategy

Search in parallel. Learn from failures. Keep iterating.

Never give up after one search round. Launch new searches based on what worked and what didn't.

Available Tools

• WebSearch - General info, news, tutorials, overviews (free, fast)
• mcp__exa__get_code_context_exa - Code examples, implementations ($0.01/query)
• mcp__exa__web_search_exa - Academic papers, technical docs ($0.01/query)

Related Skills & Snippets

• google-scholar skill - For academic paper research with citation tracking
• DOWNLOAD snippet - For downloading PDFs from Anna's Archive
• document-skills:pdf - For extracting text from downloaded papers

Iterative Workflow

Round 1: Launch Parallel Searches

Start with 3-5 searches from different angles:

General topics:

• WebSearch with variations of query
• Different phrasings, synonyms, related terms
• Add context: year "2024 2025", tech stack, keywords

Code/implementation:

• WebSearch for overview/tutorial
• mcp__exa__get_code_context_exa for real examples

Academic/research:

• WebSearch for overview
• mcp__exa__web_search_exa for papers
• Use google-scholar skill for deeper research

Round 2: Analyze & Iterate

What did you learn?

• New terminology discovered?
• Specific authors/sources mentioned?
• Related topics to explore?

What's missing?

• Gaps in understanding
• Unanswered questions
• Need more examples/evidence

Launch next round:

• Reformulate queries with new terms
• Search for specific gaps
• Try broader or narrower scope
• 3-5 new parallel searches

Round 3+: Keep Going

Don't stop until:

• Question fully answered
• Multiple sources confirm findings
• Examples found and verified
• All gaps identified and filled

If stuck:

• Try different keywords/synonyms
• Split complex query into parts
• Search related topics first
• Switch tools (WebSearch ↔ Exa)
• Check alternative sources

Query Optimization

Make queries specific:

• ❌ "authentication"
• ✅ "JWT authentication Express.js tutorial 2025"

Add context:

• Technology: "Node.js", "Python", "React"
• Timeframe: "2024 2025" for recent
• Type: "tutorial", "example", "production", "best practices"

Use domain filtering:

• Academic: allowed_domains: ["*.edu", "arxiv.org"]
• Developer: allowed_domains: ["github.com", "dev.to"]

Persistence Rules

Never give up after one failed search.

If results are poor:

1. Try different keywords
2. Broaden or narrow scope
3. Split into smaller questions
4. Switch tools
5. Search related topics first

Keep iterating until satisfied.

Common Research Patterns

Quick Answer

Single focused WebSearch with specific query

General Research

• 3-5 parallel WebSearch with different angles
• Iterate based on results

Code Research

• WebSearch for overview
• mcp__exa__get_code_context_exa for examples
• Iterate for edge cases, error handling

Academic Research

• WebSearch for overview
• mcp__exa__web_search_exa for papers
• Use google-scholar skill for citations
• Use DOWNLOAD snippet for PDFs
• Use document-skills:pdf to extract content

Example: Iterative Search

Round 1:

WebSearch: "Python async programming"
→ Learned: asyncio is the main library

Round 2:

WebSearch: "Python asyncio tutorial 2025"
WebSearch: "asyncio vs threading when to use"
mcp__exa__get_code_context_exa: "Python asyncio examples"
→ Learned: Good basics, but need error handling

Round 3:

WebSearch: "asyncio error handling best practices"
WebSearch: "asyncio common mistakes"
mcp__exa__get_code_context_exa: "asyncio exception handling"
→ Complete understanding achieved

Validation

Cross-reference findings:

• Check multiple sources
• Verify recent dates (2024-2025)
• Look for authoritative sources
• Test code examples
• Confirm consistency

Quick Reference

Default: 3-5 parallel WebSearch with different angles

For code: Add mcp__exa__get_code_context_exa

For papers: Add mcp__exa__web_search_exa or use google-scholar skill

For PDFs: Use DOWNLOAD snippet

When stuck: Reformulate, try new terms, iterate

Remember: Keep going until you find complete answers