Research
Decision Tree
code
Research request → What depth?
├─ Lookup (simple fact, quick answer) → 3-5 sources, skip to Collect + Deliver
├─ Synthesis (compare options, summarize topic) → 8-12 sources, all 5 phases
├─ Analysis (architecture decision, tech evaluation) → 12-15 sources, all 5 phases + recommendation
└─ Investigation (novel/complex, no obvious answer) → 15+ sources, all 5 phases + multiple rounds
Phases
code
Phase 1: Scope → Phase 2: Collect → Phase 3: Validate → Phase 4: Synthesize → Phase 5: Deliver
Phase 1: Scope
Before searching, define the boundaries:
- •Core question — What specifically are we trying to answer?
- •Depth level — Lookup / Synthesis / Analysis / Investigation
- •Search strategy — Keywords, year constraints, domain restrictions
- •Success criteria — What would a good answer look like?
Phase 2: Collect
Execute multi-angle searches. Don't just search once — approach the topic from multiple directions:
| Search Angle | Example Query Modifier |
|---|---|
| Factual | "what is [topic]" |
| Comparative | "[topic] vs [alternative]" |
| Technical | "[topic] architecture implementation" |
| Practical | "[topic] production experience lessons learned" |
| Recent | "[topic] 2026" or "[topic] latest" |
| Critical | "[topic] problems limitations drawbacks" |
Source diversity required:
- •Official documentation (most authoritative)
- •Official blogs and announcements
- •Independent technical analysis
- •Community discussions (GitHub issues, forums)
- •Academic papers (for foundational claims)
Parallel searches: Execute 3-5 searches simultaneously when tools support it.
Phase 3: Validate
Every claim needs triangulation:
code
Claim from Source A → Find Source B that independently confirms or contradicts
├─ Confirmed by 2+ sources → High confidence
├─ Contradicted → Note conflict, investigate why
└─ Only one source → Flag as unverified, note in output
Source chain: Official docs → Independent analysis → Community verification
URL validation: Check that every cited link is accessible before including it.
Phase 4: Synthesize
Score and weight sources using:
| Factor | Weight | High Score | Low Score |
|---|---|---|---|
| Authority | 30% | Official docs, core maintainer | Random blog, no credentials |
| Recency | 25% | Current year, last 12 months | 3+ years old |
| Specificity | 25% | Directly answers the question | Tangentially related |
| Independence | 20% | No vendor affiliation | Vendor marketing material |
Conflict resolution:
- •Prefer official documentation
- •Check publication dates (newer often supersedes)
- •Note the disagreement explicitly
- •Present both perspectives if unresolved
Phase 5: Deliver
Output structure (scale to depth):
| Depth | Output Format |
|---|---|
| Lookup | 1-2 paragraphs + sources |
| Synthesis | Sections with comparison table + sources |
| Analysis | Executive summary + detailed sections + recommendation + sources |
| Investigation | Full report with findings, methodology, limitations, sources |
Every deliverable must include:
- •Inline citations
[Source Name](url)for all claims - •Sources section at the end with all referenced URLs
- •Confidence indicators for uncertain claims
- •Date of research (findings may become outdated)
Anti-Patterns
| Anti-Pattern | Fix |
|---|---|
| Single-source claims | Triangulate with 2+ independent sources |
| Hallucinated citations | Never fabricate URLs — only cite pages you actually fetched |
| Vendor benchmarks as neutral | Explicitly note source bias |
| Overly broad queries | Focus each query on a single specific aspect |
| Missing recency | Add current year to search queries |
| Shallow exploration | Read 10+ sources, not just the first 1-2 results |
| No conflict disclosure | Always note when sources disagree |
| Broken links in report | Validate every URL before including |
| Assuming first result is best | Check multiple results, compare quality |
| Skipping the scope phase | Always define what "done" looks like before searching |