Research Synthesizer

Ten sources, three of which disagree, do not become knowledge by being summarized in sequence. Synthesis is a different operation: extract the claims, grade the evidence behind each one, put the sources in conversation — where they agree, where they collide, and why — and emerge with findings ranked by confidence plus an honest register of what remains unknown. This skill runs that operation on whatever pile of documents, papers, transcripts, or notes it is pointed at.

When to use this skill

Multiple documents, reports, or transcripts must inform one decision or one written brief
A literature scan produced twenty open tabs and someone must now say "so what?"
Two credible sources contradict each other and the difference actually matters
Assembling a state-of-knowledge brief before commissioning expensive primary research

Workflow

Fix the question before reading. Write down the two or three questions the synthesis must answer. Without them, synthesis regresses to summary — and a summary of ten sources is just a longer read.
Extract claims, not summaries. From each source, pull the specific claims relevant to the questions. Each claim carries: the source, its exact statement, the evidence type behind it — experiment, large-sample data, small-sample data, case account, expert judgment, or a claim from a party that profits by it — and its date.
Grade evidence strength per claim using anchored levels. Strong: sound method, adequate sample, disinterested source. Moderate: reasonable method with real limits. Weak: anecdote, tiny sample, or interested source. Grade the evidence, never the prestige of the publisher.
Build the agreement map. Cluster claims by question. Where independent sources converge, say so — and check independence, because three articles citing one study are one source wearing three hats. Where sources conflict, do not average them; check first whether definitions, populations, or time periods differ before declaring a genuine disagreement.
Write findings in confidence order. High-confidence findings — multiple independent strong sources — stated plainly and first. Moderate findings hedged precisely: "directionally supported by two studies, both with sample limits." Contested points presented as contested, with the strongest case for each side in one sentence apiece.
Register the unknowns. List what the questions needed that no source answered, and what evidence would settle each item. The unknowns register is often the most decision-relevant section — it is the shopping list for the next research dollar.
Cite as you go. Every finding traces to named sources inline. A synthesis whose claims cannot be walked back to sources is an opinion piece with a bibliography.

Output format

A brief with: the questions; findings in confidence order, each with supporting sources and evidence grades; a disagreements section showing both sides and the suspected cause of divergence; the unknowns register with what-would-settle-it notes; and a source table — source, type, date, interest or bias note, and the claims it contributed.

Quality bar

No claim without a citation, and no citation laundering: secondary reports traced to their primary source before they count
Independence checked; co-citing sources counted once
Conflicts explained or explicitly left open — never silently averaged, never resolved by recency alone
Interested sources admitted as evidence only with their interest labeled in the same sentence
A reader can disagree efficiently: the path from evidence to finding is visible at every step

Worked example: one finding, properly hedged

Finding (moderate confidence): "Longer onboarding sequences correlate with higher 90-day retention in the sources reviewed." Support: two independent large-sample analyses agree; one small-sample case account disagrees. Divergence note: the dissenting source measured retention at 30 days, not 90 — likely a definition difference, not a contradiction. What would settle it: any experiment randomizing sequence length, which no reviewed source ran.