deep-research-lcot

Use for deep research, literature reviews, technical due diligence, scientific explainers, market/technical reports, and reasoning-first knowledge synthesis. Turns a question into a verified research

SKILL.md

GitHub

Deep Research LCoT Skill

Purpose

Perform deep research by retrieving and synthesizing verifiable reasoning chains, not just isolated facts. Favor first-principles derivations, explicit assumptions, evidence quality, cross-domain links, and contradiction checks.

Required behavior

  1. Clarify the research objective only when a missing constraint would materially change the answer. Otherwise make reasonable assumptions and state them.
  2. Convert the user request into a research brief with: objective, scope, exclusions, decision/use case, time sensitivity, required evidence quality, and output format.
  3. Decompose the topic into atomic concepts, claims, mechanisms, prerequisites, and unknowns.
  4. Search for sources using a tiered source strategy:
    • primary sources: papers, standards, official docs, datasets, regulatory filings, laws, patents, source code
    • secondary expert sources: textbooks, review papers, technical blogs from recognized experts
    • tertiary sources only for orientation: encyclopedias, summaries, general media
  5. Build an evidence matrix. For every important claim, include source, date, author/org, evidence type, confidence, limitations, and whether it is directly supported or inferred.
  6. Build reasoning chains. Each chain must include: starting principles, assumptions, intermediate steps, endpoint claim, falsification tests, and unresolved gaps.
  7. Run adversarial verification:
    • find contradictory evidence
    • check whether sources share a common upstream source
    • distinguish empirical evidence from model/speculation
    • flag circular citations and unsupported causal claims
    • check dates for obsolete facts
  8. Synthesize with citations. Do not overstate certainty. Separate facts, estimates, interpretations, and recommendations.

Output modes

Use the mode that best fits the user request.

Research brief

Return:

  • Executive answer
  • Key findings
  • Evidence matrix
  • Reasoning chains
  • Disagreements/uncertainties
  • Practical implications
  • Next research steps

Inverse knowledge search

For a target concept, return:

  • Definition
  • Upstream prerequisite map
  • Multiple derivation paths
  • Cross-domain analogies
  • Where each path breaks
  • Best sources to verify each step

Technical due diligence

Return:

  • Claim inventory
  • Architecture or mechanism map
  • Competitive landscape
  • Evidence grade
  • Failure modes
  • Risks and mitigations
  • Decision recommendation

Citation rules

  • Cite every non-obvious factual claim.
  • Prefer direct source citations over secondhand citations.
  • Include publication or update dates when recency matters.
  • Never cite a source that does not support the sentence it is attached to.
  • Mark unsupported but plausible reasoning as inference.

Final quality checklist

Before finalizing, verify:

  • The answer addresses the user's actual decision or goal.
  • Major claims have citations or are clearly labeled as inference.
  • Dates are explicit for changing facts.
  • Contradictions and uncertainties are visible.
  • The final answer is useful, not just comprehensive.