Check Against the Source, Not the Summary

The mistake of accepting Claude's summary, quote, or paraphrase of a document as accurate because it reads well — without checking the specific claims back against the source. A fluent summary can still misstate, over-generalise, or invent a detail the original never contained.

Summarising compresses, and compression is where meaning gets bent: a hedge dropped, a 'some' turned into 'most', a figure rounded the wrong way, a caveat lost. The output reads confident and coherent, which is exactly why the error slips past — you're judging the prose, not the fidelity. The fix isn't distrust of every summary; it's checking the load-bearing claims — the numbers, quotes, and conclusions you'll actually act on — against the source itself. Verify the claim, not the polish, because polish is not evidence of accuracy.

Whenever a summary or extraction will be used for something that matters: quoting a document publicly, making a decision on its figures, relaying it to others as fact. Especially with long sources you didn't read yourself — that's precisely when you can't feel a misstatement.

Low-stakes gist-getting where approximate is fine — a rough sense of a casual article doesn't need line-by-line verification. Match the checking effort to what rides on the claim being exactly right.

Summarise this source, and make it verifiable: for each key claim, figure, or quote in your summary, cite the exact sentence or section it comes from so I can check it against the original. Flag anything you inferred rather than found stated. Source: <paste or attach>.

Claude summarises a report as 'the policy cut costs by 30%'. Checking the source, the report actually said costs fell 'up to 30% in the pilot region' — a scoped, best-case figure, not a general result. The summary wasn't lying; it dropped the qualifier that changed the meaning. You'd only have caught it by reading the claim back against the source.

Ask Claude to produce the summary as claims paired with verbatim supporting quotes, then spot-check the quotes actually appear and actually say what the claim implies. Requiring the model to anchor each claim to a quote both makes your verification fast and discourages the confident paraphrase that drifts from the text.

• Treating fluency and confidence as evidence of fidelity — a well-written summary can be subtly wrong.
• Verifying the trivial claims and waving through the one load-bearing figure your decision depends on.
• Quoting Claude's paraphrase as if it were the source's own words, when it's an approximation.