Verified against Perplexity Comet · 2026-08-10
Pull every citable claim out of a page (or thread of pages) into a table with exact quotes and provenance
A structured extraction prompt for Comet that separates verbatim quotes from paraphrase, keeps page-level provenance for each one, and flags anything that can't be traced back to where it's attributed — built for anyone assembling a sourced document from multiple browser tabs.
The prompt
Ready to copy — highlighted parts are example details you can swap.
Extract every claim from the source(s) below that would need a citation if I used it in an internal competitive-analysis memo that will be reviewed by legal before wider distribution, and build a citation table I can paste directly into that document. SOURCE(S) TO EXTRACT FROM three competitor pricing pages and one analyst report PDF, all currently open WHAT COUNTS AS CITABLE HERE pricing figures, feature availability claims, and any stated customer counts — not general marketing language CITATION FORMAT NEEDED informal internal style: [Source name, page/section, access date] EXTRACTION RULES Go through the source(s) and pull out every claim matching pricing figures, feature availability claims, and any stated customer counts — not general marketing language — do not paraphrase a claim into your own words in the extraction itself; capture it as close to verbatim as the source states it, and mark clearly whether what you captured is a direct quote (use quotation marks exactly as in the source) or your own compressed restatement of a longer passage (mark this explicitly as "paraphrased"). For every claim, note the exact page or section it came from — not just the site name — so I can find it again without re-reading the whole source. If a claim on the page is itself attributed to another source that isn't the page in front of you (e.g. "according to a 2025 industry report..."), note that the citation trail goes one level deeper and flag it rather than citing the page you're reading as if it were the origin. OUTPUT TABLE Claim (quote or paraphrase, labeled) | Exact location (page/section/paragraph) | Attributed origin if different from the page itself | Formatted citation in informal internal style: [Source name, page/section, access date]. WHAT NOT TO DO Do not invent a page number, publication date, or author name if the source doesn't actually display one — write "not stated on page" rather than guessing, since a fabricated citation detail is worse than an honest gap. Do not silently merge two similar claims from different parts of the source into one row if they're phrased differently — list them separately even if redundant, and let me decide whether to consolidate. End with a short list of any claims you found compelling but could not locate a specific, quotable location for — these need a second look before they go in the document.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
The quote-versus-paraphrase distinction is the single highest-leverage rule in this prompt because a language model's default extraction behavior is to lightly rephrase everything into consistent prose, which quietly turns an exact quote into something that looks like one but isn't — a difference that matters enormously the moment a citation gets checked against its source. Requiring page- or section-level location rather than just the site name matters because Comet is browsing live pages, not summarizing an index entry, so it actually has access to where on the page a claim sits — asking for that level of detail uses the one capability a browsing agent has that a plain search-summary tool doesn't. The instruction to flag one-level-deeper attribution (a page citing another source) exists because extraction tools commonly collapse a chain of attribution into a single citation to the page in front of them, which silently misattributes a claim to whoever happened to repeat it rather than whoever originated it — flagging this explicitly keeps the citation trail honest instead of one hop short. The explicit ban on inventing missing details (a page number, an author, a date) addresses a known model failure mode under table-formatting pressure: an empty cell feels incomplete, and the path of least resistance is to fill it with something plausible rather than admit the source doesn't provide it — writing 'not stated on page' as the required fallback removes the incentive to guess. Keeping duplicate-seeming claims as separate rows rather than letting the model silently consolidate them preserves editorial control over a judgment call that belongs to the person assembling the document, not to the extraction step.
What you get back
Claim: "pricing starts at $49/seat/month for teams under 20" (direct quote) | Location: pricing page, second pricing tier card | Attributed origin: the page itself | Citation: [Competitor X pricing page, 2026-08-09]. Claim: over 10,000 companies use the product (paraphrased from "joins the 10k+ companies already...") | Location: homepage hero section | Attributed origin: page itself, no methodology given for the count | Citation: [Competitor X homepage, 2026-08-09]. Flagged for second look: a claim about "industry-leading uptime" with no specific figure or source named anywhere on the page.
Verified against
Perplexity Comet 2026 browser agent · 2026-08-10
Changelog
- 2026-08-10 — Initial publish, verified against Perplexity Comet 2026 browser agent.
Need this built into your business?
If a prompt isn't enough — what Scult builds, built and maintained for you — that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
