The Research Synthesizer: Turn Scattered Sources Into a Map of Agreements and Contradictions
Why this prompt matters
Without a structured synthesis step, people default to citing whichever single source already matched what they expected to find, and the gap surfaces at the worst possible moment — when a stakeholder points out mid-meeting that another source in the same pile actually argues the opposite. That credibility hit costs far more time and trust to recover from than the twenty minutes a proper synthesis would have taken upfront.
What we use it for
A product manager has four analyst reports and a dozen saved articles on whether on-device AI or cloud AI will dominate enterprise deployment by 2027, needs to present a one-page synthesis to the leadership team by Friday afternoon, and does not have time to read everything closely enough to defend a recommendation under questioning in the room.
Prompt
Role: Act as a research synthesis analyst who specializes in finding genuine points of agreement, disagreement, and unaddressed gaps across multiple sources covering the same question — not a summarizer who blends everything into one smooth narrative. Context: I am researching [YOUR RESEARCH QUESTION]. Here are the sources, each clearly labeled: Source A: [PASTE SUMMARY, ABSTRACT, OR FULL TEXT] Source B: [PASTE SUMMARY, ABSTRACT, OR FULL TEXT] Source C: [PASTE SUMMARY, ABSTRACT, OR FULL TEXT] [ADD MORE SOURCES AS NEEDED, LABELED SOURCE D, E, ETC.] Task: Produce a structured synthesis that: 1. Identifies the 3 to 5 core claims being made across all sources combined 2. For each core claim, states which sources agree, which disagree, and quotes or closely paraphrases the specific language showing the agreement or disagreement 3. Distinguishes disagreements that are factual (verifiable, at least one source is likely wrong) from disagreements that are interpretive (both sources can be reasonably right, they emphasize different things) 4. Identifies at least [NUMBER] unaddressed gaps — specific sub-questions none of the sources cover, that would need separate research 5. Flags any source that appears to have a conflict of interest or vested stake in a particular conclusion, such as a vendor-sponsored study or an advocacy group Constraints: - Do not average disagreements into a vague middle-ground conclusion. State disagreements explicitly, with attribution to the specific source. - Do not treat source volume as a proxy for correctness. Three sources repeating the same unverified claim is not stronger evidence than one source citing primary data — check whether sources are independently reporting or just citing each other before counting them as separate confirmations. - If two sources make the same claim but one cites a primary source and the other cites the first source, note this explicitly as one underlying data point, not two. - Keep the total output under [WORD LIMIT] words. Output Format: 1. A short summary paragraph (3 to 4 sentences) stating the overall state of consensus or lack of it 2. "Where Sources Agree" — list of claims with supporting sources 3. "Where Sources Disagree" — list of claims with which sources say what, and whether the disagreement is factual or interpretive 4. "Unaddressed Gaps" — list of the open sub-questions 5. "Caveats" — short section flagging any source credibility concerns
Result
Overall State of Consensus: Of the three sources reviewed, there is partial agreement that on-device inference will handle a growing share of routine enterprise AI tasks by 2027, but the sources sharply disagree on whether this displaces cloud AI for core workloads or simply adds a new tier alongside it. Source C's central claim appears to rest heavily on Source A's original survey data without independent verification.
Where Sources Agree: - On-device inference costs have dropped enough to handle routine classification and short-context tasks locally (Sources A, B, C) - Enterprise IT buyers cite data residency, not raw cost savings, as the top driver for on-device adoption (Sources A, B)
Where Sources Disagree: - Claim: "On-device AI will capture the majority of enterprise AI spend by 2027." Source A projects 35% share by 2027 based on a survey of 200 IT buyers; Source B projects under 15% share, arguing complex reasoning tasks still require cloud-scale models regardless of hardware improvements. This is a factual disagreement — the two sources make incompatible quantitative predictions, and only one survey methodology can be sounder. Worth checking sample composition before trusting either figure. - Claim: "Cloud AI providers are cutting prices industry-wide to defend market share." Source C asserts this is already happening broadly; Source A documents only one vendor doing this. This is an interpretive disagreement — Source C may be generalizing from a single data point rather than an industry pattern.
Unaddressed Gaps: - No source addresses how data-residency requirements outside the US and EU affect this split - No source quantifies the latency threshold at which enterprises actually migrate workloads from cloud to on-device, versus simply running new workloads on-device from the start
Caveats: Source C is published by a chipmaker with a direct commercial interest in on-device AI adoption and should be weighted accordingly — its 35%-by-2027 figure traces back to Source A's survey rather than independent data.
The usual way people handle conflicting sources is to pick the one that already matches what they wanted to believe, or to ask an AI tool for a quick summary that blends everything into a single confident paragraph. Both approaches quietly erase the most useful information in the pile: the specific points where credible sources actually disagree, and why.
Why the Role matters here
The Role frames the model as a synthesis analyst rather than a summarizer, because those are different jobs. A summarizer's job is to compress; a synthesis analyst's job is to map disagreement and attribute it correctly. Asking a general-purpose assistant to “summarize these four reports” almost always produces the smoothed-over version, because smoothing is what summarization optimizes for by default.
What the Task section is actually doing
The five numbered requirements force a specific sequence: find the core claims first, then check agreement per claim, then classify disagreements as factual versus interpretive. That last distinction is the one most people skip, and it is the one that matters most. A factual disagreement means someone is probably wrong and it is worth digging into methodology. An interpretive disagreement means both sources can be right while emphasizing different things — treating it as a factual dispute wastes time looking for a “correct” answer that does not exist.
The Constraint that does the real work
“Do not average disagreements into a vague middle-ground conclusion” is the instruction protecting the entire point of the exercise. Left unconstrained, most models will resolve tension by finding soft, hedged language that technically accommodates every source — which is exactly the output that gets someone caught flat-footed when a colleague points out the sources never actually agreed. The citation-counting constraint matters just as much: three blog posts that all cite the same original study are one data point wearing three outfits, not three independent confirmations.
Why the output format is split into four sections
Separating “Where Sources Agree” from “Where Sources Disagree” forces the disagreements to stay visible instead of disappearing into a combined narrative. The “Unaddressed Gaps” section exists because the most expensive research mistake is not misreading a source — it is assuming a question was covered when it was never asked. The “Caveats” section catches the sources that have a reason to prefer one conclusion, which matters more than source count when deciding how much weight to give a claim.
Run this before any meeting, memo, or decision that depends on reconciling several sources that disagree with each other — not after, when the gaps have already become someone else's objection.