research-synthesis

name	research-synthesis
description	Synthesise findings across multiple documents or web sources into a structured research report with inline citations, contradiction detection, and evidence grading. Use when asked to "research a topic across documents", "synthesise findings", "literature review", "compare sources", "cross-reference documents", "compile research report", "what do these documents say about X", or any multi-source analysis task.
license	MIT
metadata	{"version":"1.0.0","tags":["research"]}

Multi-Document Research Synthesis

When to Use This Skill

Use this Skill when the user asks to:

• Research a topic across multiple documents or sources
• Synthesise or reconcile findings from different sources
• Produce a literature review or research summary
• Detect contradictions between sources
• Grade evidence quality across a corpus
• Answer a question using multiple documents as evidence
• Compile a cross-referenced research report

Workflow

Research Synthesis Progress:

• Step 1: Define research question and gather sources
• Step 2: Extract and chunk all sources
• Step 3: Retrieve relevant passages
• Step 4: Analyse and cross-reference
• Step 5: Detect contradictions
• Step 6: Grade evidence
• Step 7: Synthesise findings
• Step 8: Produce report
• Step 9: Persist output

Step 1: Define research question and gather sources

Clarify:

• Research question — What specifically are we investigating?
• Source corpus — Documents in storage (local files), uploaded files, web URLs, or search index?
• Scope constraints — Date range, source types, specific authors/publications?

Gather sources: Stored documents → Tool: local_files({ path }) Web sources → Tool: browse_web({ url }) Search index → Tool: query_search_index({ query, top_k: 50 })

Step 2: Extract and chunk all sources

For each source:

• PDF → Tool: extract_pdf({ source }) or call_doc_intelligence({ source })
• DOCX → Tool: extract_docx({ source })
• XLSX → Tool: extract_xlsx({ source })
• Web → Already extracted via browse_web

Chunk each source into passages (~300-500 tokens) preserving:

• Source attribution (file name, URL, page number)
• Section context (heading hierarchy)
• Passage index within source

Step 3: Retrieve relevant passages

For each aspect of the research question: Build a relevance matrix: which passages from which sources address which aspects of the question.

Step 4: Analyse and cross-reference

For each relevant passage:

• Extract the claim (what the source asserts)
• Identify supporting evidence (data, citations, reasoning)
• Note the context (methodology, assumptions, caveats)
• Tag the stance (supports, contradicts, neutral, or qualifies the research question)

Cross-reference claims across sources:

• Corroboration — Multiple sources support the same claim
• Contradiction — Sources make incompatible claims
• Complementarity — Sources address different aspects without conflict
• Qualification — One source adds conditions or nuance to another's claim

Step 5: Detect contradictions

For each identified contradiction:

Source A	Claim A	Source B	Claim B	Nature	Resolution
				Direct / Partial / Contextual	Which is more credible and why

Contradiction types:

• Direct — Factual claims that cannot both be true
• Partial — Claims that overlap but differ in scope or magnitude
• Contextual — Claims that appear contradictory but apply to different contexts

Step 6: Grade evidence

Score each source and claim using the evidence grading rubric:

Grade	Label	Criteria
A	Strong	Primary data, peer-reviewed, large sample, reproducible methodology
B	Moderate	Reputable secondary source, expert analysis, consistent with other evidence
C	Weak	Single source, opinion-based, small sample, potential bias, outdated
D	Very Weak	Anecdotal, unverified, known bias, contradicted by stronger evidence
U	Unknown	Insufficient information to assess credibility

Factors:

• Source authority (expertise, publication venue, institutional backing)
• Methodology (if applicable — sample size, controls, replicability)
• Recency (newer data on fast-moving topics scores higher)
• Corroboration (claims supported by multiple independent sources score higher)
• Potential bias (funding source, advocacy position, commercial interest)

Step 7: Synthesise findings

Construct a narrative synthesis that:

1. Opens with the consensus view — what most/strongest sources agree on
2. Addresses key debates — where sources disagree, with evidence for each side
3. Identifies gaps — aspects of the question no source adequately addresses
4. Draws conclusions — qualified by evidence strength
5. Uses inline citations — [Source Name, p.X] or [URL, §Section] for every claim

Step 8: Produce report

Research Summary

Field	Value
Research question
Sources analysed
Date range
Evidence quality	(A/B/C distribution)

Key Findings

1. [Finding] — [Evidence summary with citations] — Evidence grade: [A/B/C]

Contradictions & Debates

Topic	Position A	Position B	Stronger Evidence	Resolution

Evidence Gap Analysis

Aspect	Coverage	Recommendation

Source Assessment

Source	Type	Recency	Authority	Relevance	Evidence Grade

Narrative Synthesis

[Multi-paragraph synthesis with inline citations]

Bibliography

#	Source	Type	Date	URL/Location

Step 9: Persist output

Tool: local_files({ path: "~/contelligence-output/research//.md", content: })

Edge Cases

• Sources in different languages: Note the language of each source; synthesise in the user's requested language.
• Highly technical sources: Flag domain-specific jargon and provide brief definitions.
• Sources with different publication dates: Weight recent sources higher for fast-evolving topics; note temporal context for each claim.
• Very large corpus (>20 sources): Prioritise by relevance score; provide a tiered report (top findings from all sources, detailed analysis of top 10).

Constraints

• Every claim in the synthesis must have at least one citation — never present unsourced assertions.
• Clearly distinguish between what sources say and your own analytical conclusions.
• When evidence is weak or conflicting, state the uncertainty explicitly.
• Do not cherry-pick — present the full range of perspectives found in the sources.