Prompt: Summarise and Compare Sources

A prompt for comparing multiple sources that surfaces where they actually agree, disagree, and simply don't cover the same ground.

The prompt

Open in Studio

You are a research assistant whose job is to show where sources agree and disagree, not blend them into one voice.

Objective: Compare the supplied sources on the question: {{RESEARCH QUESTION}}.

Context: The sources are {{BRIEFLY DESCRIBE THE SOURCES — e.g. three vendor comparison articles, two academic papers}}.

Constraints:
- State which specific source each claim comes from
- Distinguish claims the sources agree on from ones where they conflict
- Note anything the question asks about that none of the sources actually cover
- Do not resolve a conflict between sources yourself — present it as unresolved

Example: if Source A states a 20% cost reduction and Source B doesn't mention cost at all, that's a coverage gap, not a disagreement — don't conflate the two.

Output format: Return three sections — Where Sources Agree, Where Sources Disagree, and Not Covered by Any Source — each as a Markdown list with the source cited per point.

Before finishing, check that every claim is attributed to a specific source and that no source is silently left out.

Sources:
{{PASTE SOURCE TEXTS OR SUMMARIES HERE}}

Example

Input

RESEARCH QUESTION: does remote work increase or decrease team productivity?
BRIEFLY DESCRIBE THE SOURCES — e.g. three vendor comparison articles, two academic papers: two workplace research articles
PASTE SOURCE TEXTS OR SUMMARIES HERE:
  Source A: a 2023 internal study at a software company found that fully remote teams shipped features 15% faster than co-located teams, attributing this to fewer meetings and more uninterrupted focus time.
  Source B: a survey of managers across several industries found that remote workers reported lower feelings of team cohesion, though the survey did not measure output or delivery speed directly.

Output

Where Sources Agree
- No shared claim to agree on: Source A measures delivery speed (features shipped 15% faster by fully remote teams), while Source B measures self-reported team cohesion (lower among remote workers).

Where Sources Disagree
- No direct disagreement: the two sources measure different outcomes, so Source A's faster delivery and Source B's lower cohesion don't contradict each other (Source A, Source B).

Not Covered by Any Source
- Overall team productivity, which is what the question asks about: Source A covers delivery speed at one software company only, and Source B states that its survey did not measure output or delivery speed directly.

- Whether faster delivery (Source A) and lower cohesion (Source B) occur in the same teams.

Why it works

Requiring a named source per claim makes it possible to check the model's synthesis against what a source actually said.
Separating agreement, disagreement and coverage gaps stops a coverage gap being mistaken for a genuine disagreement, a common conflation.
Banning the model from resolving a conflict itself keeps a genuine disagreement visible instead of silently smoothed into one voice.

Why the basic version fails

“Summarise these sources” tends to produce one smooth synthesis that quietly picks a side in a genuine disagreement, or drops the source that didn’t fit the narrative — without ever telling you it did that.

What this prompt fixes

Requiring per-claim attribution makes it possible to check the model’s work against the actual sources. Explicitly asking for disagreement and coverage gaps, rather than a single blended answer, surfaces exactly the information a synthesis would otherwise hide.

Tips per model

Claude handles multiple long sources well and tends to keep attribution consistent throughout.
Gemini is a reasonable choice for a larger number of sources; spot-check a few attributed claims against the originals before relying on the comparison.
Label each source clearly (Source A, Source B, and so on) before pasting them in — consistent labels make the per-claim attribution in the output far easier to verify against the originals.