Deep Research Prompt

Deep Research Prompt.md

<ResearchParameters>
  <Topic>[Insert Specific Research Topic Here]</Topic>
  
  <KeyQuestions>
    <!-- Optional: List 1-5 specific questions you want answered -->
    <!-- Example: What are the primary causes of X? -->
    <!-- Example: Compare and contrast approaches Y and Z. -->
    <!-- Example: What is the projected market size for Q by 2030? -->
    <!-- If left empty, the model will perform a broad investigation -->
   </KeyQuestions>
  
  <SuppliedSources>
    <!-- Optional: paste authoritative docs, specs, papers here -->
    <!-- When populated, model must prefer these over external search -->
  </SuppliedSources>

  <ReportRequirements>
    
    <!-- EPISTEMIC STRUCTURE (Required) -->
    For each major claim or finding, explicitly separate:
    
    1. **Definitions & Prerequisites**
       - Define key terms precisely; note competing definitions if they exist
       - State what background knowledge is assumed
       
    2. **Claims** — The actual assertion being made
    
    3. **Evidence** — One of:
       - Direct quote from source (with citation)
       - Data point with provenance
       - Derivation with explicit steps
       - If from <SuppliedSources>, quote exact lines relied upon
       
    4. **Assumptions** — What must be true for this claim to hold
       - If assumptions are unknown or contested, flag them
       
    5. **Confidence** — For each key claim:
       - Estimated confidence as [low / medium / high] with brief rationale
       - What evidence would raise or lower confidence
       
    6. **Failure Modes** — Where could this be wrong? What would falsify it?
    
    <!-- VERIFIABILITY HOOKS (At least one per major claim) -->
    Every significant claim must include at least one:
    - Pointer to primary source (docs, paper, standard) with specific section/page
    - Minimal working example (if technical)
    - Testable prediction or counterexample search
    - Quote from <SuppliedSources> if provided
    
    If no verifiable hook exists, explicitly label the claim as **[HYPOTHESIS - UNVERIFIED]**.
    
    <!-- SOURCE DISCIPLINE -->
    - When <SuppliedSources> is populated: use only those sources; say "not in provided material" if information is missing
    - When searching externally: prioritize academic studies, institutional reports, technical papers, official documentation
    - Reject or flag: non-expert blogs, marketing content, unsourced claims
    - For any claim where sources conflict, note the disagreement explicitly
    
    <!-- DEPENDENCY GRAPH -->
    At the report's end, include a section:
    **Core Dependencies** — The 3–7 foundational claims the analysis rests on
    For each:
    - The claim itself
    - How to verify it independently
    - What breaks if this claim is false
    
    <!-- CONTENT SCOPE -->
    - Directly address all <KeyQuestions>
    - Synthesize into cohesive analysis (not just source summaries)
    - Identify trends, challenges, limitations, future implications
    - Reason from first principles where applicable
    - Be concise; prefer derivations and invariants over narrative prose for technical content
    
    <!-- SEARCH STRATEGY (Show Your Work) -->
    Before the main report, include a brief:
    - Keywords, synonyms, and related terms used
    - Source selection rationale
    - What you searched for but couldn't find
    
    <!-- STRUCTURE -->
    - Clear headings/subheadings
    - Domain-appropriate terminology for expert audience
    - Logical flow with transparent reasoning
    
  </ReportRequirements>
  
  <OutputFormat>
    1. **Search Strategy** (brief)
    2. **Executive Summary** (with confidence levels for key conclusions)
    3. **Main Analysis** (structured per epistemic requirements above)
    4. **Core Dependency Graph**
    5. **Open Questions & Verification Checklist** — What should the reader verify first?
  </OutputFormat>
  
</ResearchParameters>

Generate a research report adhering strictly to <ReportRequirements> and <OutputFormat>. 
Do not blend certainty with speculation. 
If you cannot verify something, say so explicitly.

When stakes are high, 2nd pass validation

<CriticPass>
  Review the above report. For each major claim:
  1. Identify any ambiguous terms left undefined
  2. Attempt to construct a counterexample
  3. Flag any reasoning steps that don't follow
  4. Suggest one specific verification test
  5. Note if confidence seems miscalibrated (too high/low for evidence provided)
</CriticPass>