Deep Research — Multi-Source Cross-Referencing Methodology

When to use

Use when the user asks for deep analysis of a concept, tool, claim, trend, video, article, or topic that benefits from multiple perspectives. NOT for simple fact-finding (use web_search directly for that).

Triggers: “analyze in depth”, “cross-reference”, “what’s the best approach”, “user data + expert data + evidence”, “deep research”, “crème de la crème”

Core Methodology — The 3-Axis Framework

Every deep research task must cross-reference three distinct data sources before synthesizing:

Axis 1: 🧑‍💻 Community / User Data

• What: Actual users — what works in practice, frustrations, workarounds
• Sources: HN (news.ycombinator.com via Algolia API), Reddit (subreddits), GitHub Issues/Discussions, TikTok, Twitter/X
• Signal: High when multiple independent users agree. Low for single anecdotes or hype-driven platforms (TikTok).
• Pitfall: Vocal minorities and survivorship bias. Cross-check complaints against actual usage patterns.

Axis 2: 🎓 Expert / Official Data

• What: Recognized experts, official docs, thought leaders
• Sources: Engineering blogs (Anthropic, OpenAI, etc.), Simon Willison, recognized community figures, peer-reviewed papers, official documentation
• Signal: Higher bias toward “ideal world” — experts may undersell real friction
• Pitfall: Can lag behind community reality. Check publication dates.

Axis 3: 📊 Evidence-Based / Measured Data

• What: Objective, reproducible, quantifiable
• Sources: Security audits (Snyk, Trail of Bits), benchmarks, token cost measurements, star counts, adoption stats, performance comparisons
• Signal: Highest confidence when independently verifiable
• Pitfall: Measurements age fast (3+ months = stale for rapidly evolving tools)

Workflow

Phase 1: Extract Source Material

• YouTube: oEmbed API (always works, no auth) + web_search for description/summary
• Short articles/posts: web_extract (when backend supports it) or curl -sL
• Long content (papers, docs): curl + text extraction, chunking
• Capture: Title, author, key claims with direct quotes, exact URL

Phase 2: Gather All 3 Axes (use delegate_task for parallel)

• Community: Search for concept + “HN”, “reddit”, “discussion”, “review”, “worth it”
• Expert: Search for concept + “guide”, “best practices”, “[expert name] [concept]”
• Evidence: Search for concept + “benchmark”, “security”, “tested”, “token cost”, “comparison”, “study”

Phase 3: Cross-Reference

Phase 4: Structured Synthesis

Output a hierarchical ranked synthesis — not a flat list:

## [Concept] Analysis

**Source analyzed:** [link]

### 🧑‍💻 Community says
- [bullet points]

### 🎓 Experts say
- [bullet points]

### 📊 Evidence shows
- [bullet points]

### 🔬 Cross-referenced verdict

**🥇 Best approach:**
Clear winner + reasoning

**🥈 Good alternative:**
When the winner doesn't fit

**🚫 Avoid:**
What doesn't work / overhyped

**Key pitfalls:**
- [numbered bullets]

Tips & Pitfalls (learned from experience)

• Do NOT trust community alone — hype cycles and vocal minorities are real (TikTok especially inflates tool popularity)
• Do NOT trust expert data alone — experts have a rosier view than practitioners
• Evidence beats opinion for objective claims (speed, security, cost). Community beats experts for subjective claims (DX, pleasantness, ecosystem feel).
• Token-conscious research: Extract only the most relevant signals from each source. Full-page dumps waste context.
• 🔋 TOKEN ECONOMY (user preference): This user values token efficiency above exhaustive coverage. Prefer 3-5 targeted searches over 10 broad ones. Extract only the highest-signal quotes and data points. Never dump full pages. When token budget is tight (< 5K remaining), skip community sentiment entirely and focus on evidence + expert data.
• For fast-moving topics: Filter search to last 30 days. Tools evolve weekly.
• Parallel gathering: Use delegate_task with 3 parallel agents (one per axis) when the topic is large.
• When axes strongly disagree: Note this prominently in the verdict. The tension itself is valuable information.
• Telegram output: No markdown tables. Use bullet lists and hierarchy sections.

Bridging from YouTube Shorts (no transcript → deep research)

When the source is a YouTube Short (no transcript available) and the user wants deep analysis:

1. oEmbed → extract title + channel: curl "https://www.youtube.com/oembed?url=https://youtube.com/shorts/VIDEO_ID&format=json"
2. Extract concept from title — the title is your research target. Mine it for keywords: key idea, subject, implied question.
3. Trigger deep research on the concept (not the video — Shorts have no content beyond the title). The concept IS the source material.
4. Community data: Use HN Algolia API for structured comments (curl "https://hn.algolia.com/api/v1/items/ID") — it’s reliable and not blocked (unlike Reddit JSON which blocks server IPs)
5. Cross-reference the concept normally (3-axis). The Short is just the trigger, not the content.
6. Structure output as ranked synthesis (🥇 → 🥈 → 🚫). Not a flat list.

Reference Files

• references/claude-code-ecosystem-research.md — worked example of this methodology applied to the Claude Code Skills/MCP/Plugins ecosystem
• references/cross-ref-scoring.md — how to weight and score the three axes when they disagree
• references/token-optimization.md — token economy best practices for research workflows