Deep Research
Thinking Protocol
Before starting, answer silently:
- What is the actual question? (Restate precisely — don't drift)
- What would constitute a strong vs weak answer?
- What are my known unknowns?
Execution
Phase 1: Decompose
Break into 3-7 precise sub-questions. For each, state what evidence type would answer it.
Phase 2: Multi-Source Search
Per sub-question, search minimum 3 independent sources. Prefer:
- Primary over secondary sources
- Recent over dated (flag anything >2 years old)
- Peer-reviewed/official over blog posts
- Contradicting sources are MORE valuable — seek them out
Phase 3: Evidence Grading
Every claim gets a confidence tier:
- HIGH — Multiple independent sources agree, empirical evidence
- MEDIUM — Single credible source, or sources with caveats
- LOW — Anecdotal, single blog post, or inference
- CONTESTED — Sources directly contradict (present both sides)
Phase 4: Synthesize
## [Topic]: Key Findings
### [Sub-question 1]
[Finding] — Confidence: [tier]
Evidence: [citations]
### Contradictions & Open Questions
[Where sources disagree, what remains unknown]
### Implications
[What this means for the user's actual goal]
Rules
🚨 Never present inference as fact. If reasoning beyond evidence, say so. 🚨 If confirming expectations, stop and search for disconfirming evidence. 🚨 State what you DON'T know with the same confidence you state what you do.