Extract Claims With Citations You Can Check

A summarising pattern where you ask Claude to return each key point paired with the exact quote or section it came from, so every claim in the summary is traceable back to the source. The output isn't just a digest — it's a digest you can audit line by line.

The risk in any summary is a claim that sounds right but isn't in the source, or that drifts from what the source actually said. Requiring a supporting quote for each point attacks this two ways: it makes verification fast — you check the quote instead of re-reading the document — and it discourages the confident paraphrase in the first place, because a claim with no real supporting passage has nowhere to hide. You get a summary whose trustworthiness you can spot-check in seconds rather than take on faith.

Whenever accuracy is non-negotiable: summarising research, contracts, financial or legal documents, anything you'll quote, cite, or make a decision on. Especially for long sources you can't hold in your head, where an unsupported claim is otherwise invisible.

Casual reading or brainstorming where an approximate gist is fine and citation overhead would slow you down for no benefit. Reserve the rigour for claims that have to be exactly right.

Extract the key claims from this source as a list. For each claim, give: (1) the claim in one line, (2) the exact sentence or short quote from the source that supports it, and (3) its location (section/page). If a point is your inference rather than something the source states, label it [inferred]. Source: <paste or attach>.

Summarising a 40-page report, Claude returns 'Revenue grew 12% — "...revenue rose 12% year over year in Q3" (p.14)'. You jump to page 14, confirm it in seconds, and move on. One claim it marked [inferred] — that the trend would continue — you now know to treat as its extrapolation, not the report's finding. The citations made both the solid claims and the soft one legible.

After getting the cited claims, spot-check a few quotes against the source — do they exist, and do they actually say what the claim implies? If the quotes hold, your confidence in the rest is well-founded; if one is misquoted or over-read, that's your signal to check the others. This turns verification from re-reading the whole document into sampling a few anchors.

• Accepting cited claims without ever spot-checking a quote — a citation you never verify is just decoration.
• Not distinguishing what the source states from what Claude inferred, so an extrapolation reads as a finding.
• Requesting citations but for a source Claude can't actually see (a paraphrase from memory), where the 'quotes' may be fabricated.