Does this stop hallucination?

It does not prevent it, it exposes it. A fabricated quote fails a search of the source in seconds, which is the whole value: verification becomes cheap enough that you actually do it.

What if the document is too long for one prompt?

Summarise it in chunks with the same format, then merge the numbered claims yourself. Do not summarise the summaries — that is the step where the quotes get dropped.

Docs & research

Summarise with receipts

Summarise a long PDF, paper or web page so that every claim carries the quote and location it came from — turning an unverifiable summary into a checkable one.

Use this whenever you will act on the summary, cite it, or repeat it to somebody who trusts you.

The pattern

Copy this and replace everything in [SQUARE BRACKETS] with your own specifics.

Prompt template

Below is [WHAT THE DOCUMENT IS, and how long].

[PASTE THE DOCUMENT, or attach it.]

Summarise it for [WHO, and what they need it for].

Format: [5-8] numbered claims. After each claim, on its own line:
  Source: "[the exact sentence from the document that supports it]" — [page / section / heading]

Rules:
- If you cannot quote a sentence that supports a claim, do not make the claim.
- Mark anything the document implies but never states as INFERRED, and say what it is inferred from.
- If the document contradicts itself, give both quotes and label it CONFLICT rather than picking one.
- Do not add background knowledge from outside this document.

Why each part is there

What they need it for: A summary for a decision keeps different sentences than a summary for an exam. Without this, you get the document's own emphasis rather than yours.
The Source line: The mechanism of the whole pattern. A claim with a quote attached takes ten seconds to verify; a claim without one takes a re-read of the source.
No quote, no claim: Converts the default failure — a fluent sentence with nothing behind it — into a visible omission you can go and check yourself.
INFERRED and CONFLICT labels: Summaries flatten uncertainty into confidence. These two labels are where the honesty lives: they preserve the difference between what a document says and what you wish it said.

Before and after

The lazy version

Summarise this 40-page report in ten bullet points.

What goes wrong: Ten fluent, confident bullets. Some are in the report, one is from the model's general knowledge, and one is a plausible-sounding number that appears nowhere — and there is no way to tell which is which without reading all 40 pages.

The pattern, filled in

Worked example

Below is a 34-page government report on school internet access, PDF.

[document attached]

Summarise it for a teacher deciding whether to run a computer-lab-based course next term.

Format: 6 numbered claims. After each claim, on its own line:
  Source: "[the exact sentence from the document that supports it]" — [page / section]

Rules:
- If you cannot quote a sentence that supports a claim, do not make the claim.
- Mark anything the document implies but never states as INFERRED, and say what it is inferred from.
- If the document contradicts itself, give both quotes and label it CONFLICT.
- Do not add background knowledge from outside this document.

What that buys you: Six claims, six quotes with page numbers, one CONFLICT where the summary tables and the appendix disagree on connectivity rates — the single most decision-relevant thing in the document, and exactly the thing a normal summary would have quietly averaged away.

Now do it on your own work

Take a real document over 10 pages — a chapter, a report, a paper. Run the pattern, then verify the quotes yourself by searching the source for each one. Count the hits and the misses.

Check yourself — every box should be true

You checked every Source quote against the document by searching for it.
You found and recorded any quote that was paraphrased rather than exact.
At least one claim was labelled INFERRED or CONFLICT, or you can confirm the document was genuinely consistent.
You can state one thing the summary left out that mattered for your purpose.

Then explain it back

Requiring a quote changed what the summary was willing to claim. What does that tell you about the confident-sounding claims in an ordinary summary?

Write two or three sentences in your own words. If you cannot, you copied a prompt rather than learning a pattern — which is the whole difference this library is trying to make.

Does this depend on which AI tool I use?

Two real differences here. First, page numbers: tools that read the PDF natively can usually cite pages, while tools that receive pasted text cannot — ask for the nearest heading instead. Second, quote fidelity: some tools paraphrase inside quotation marks, so spot-check by searching the source for the exact string. Trust the tool only as far as your spot-checks hold.

Why it actually works

You just built the core of a retrieval-grounded system by hand: an answer is only as good as the passage it can point at.

How retrieval grounding actually works →

Prompting well is a skill you can pick up in a week. Understanding what is happening underneath is the 18-month curriculum — and it is what stays useful when the tools change.

FAQ

Does this stop hallucination?: It does not prevent it, it exposes it. A fabricated quote fails a search of the source in seconds, which is the whole value: verification becomes cheap enough that you actually do it.
What if the document is too long for one prompt?: Summarise it in chunks with the same format, then merge the numbered claims yourself. Do not summarise the summaries — that is the step where the quotes get dropped.

Next patterns

The Prompt Patterns Library is a separate, optional track — it does not change the 78-week curriculum or your roadmap. See what the curriculum actually covers →