How to summarise long documents with ChatGPT

Check it read everything first

Before you judge the summary, establish what it is a summary of.

Documents past 2 million tokens get truncated, and ChatGPT does not mention it. You get a confident summary of the first portion that reads exactly like a summary of the whole thing.

The test takes ten seconds:

Before summarising: what is the title of the final section, and what does the document's last paragraph say?

If it cannot answer, everything that follows is based on part of the file. Split the document and send it in parts, or send only the section you need. There is more on that in getting past the upload limit.

Say who it is for

The single biggest improvement available, and it costs one sentence.

"Summarise this document" has no audience, so you get the blandest possible reading. Compare:

Summarise this for a finance director who needs to decide whether to renew the contract. Focus on cost changes, termination terms, and anything that increases our liability. Under 300 words.

Same document, completely different output. You have told it what matters and what to leave out, which is most of what summarising actually is.

Prompts that produce different shapes

Pick the shape that fits what you are going to do with it.

For a decision:

Summarise this and end with the three things I need to decide, each with the evidence from the document that bears on it.

For a meeting:

Give me five bullets I can read out in a meeting. One line each, plain language, no jargon.

For working through a long report:

Summarise section by section, using the document's own headings. Two or three sentences per section. Do not merge sections together.

For finding the disagreements:

What does this document claim that is not supported by evidence in it, and where does it contradict itself?

That last one is the most underrated. Everyone asks what a document says. Asking where it is weak tells you something you could not have got by skimming.

Getting quotes rather than paraphrase

The fastest way to make a summary verifiable:

For each point, quote the sentence it came from and give the section it appears in.

Two benefits. You can check the claim without rereading, and paraphrase drift mostly disappears, because a quote either exists or it does not.

Summarising several documents at once

The mistake is uploading all of them and asking for one summary. What comes back is mush, because the sources have been blended and you cannot tell what came from where.

Do it in two passes instead. Summarise each document separately, then paste those summaries back in and ask the comparative question:

Here are summaries of four vendor proposals. Where do they agree, where do they differ materially, and which one has the weakest evidence for its claims?

Slower, and much better. It also keeps attribution intact, which matters the moment someone asks where a figure came from.

Verify anything numeric

Summarising is where numbers go wrong. Figures get rounded, attached to the wrong year, or blended between two tables.

Ask for the source sentence behind each number, then spot-check two or three against the document. If those hold, the rest usually does. If one is wrong, check them all.

A workflow that holds up

1. Confirm the whole document was read, using the last-section test.
2. Say who the summary is for and what it supports.
3. Ask for a specific shape and a specific length.
4. Require a quote and a section reference for each point.
5. Verify the numbers.

The whole thing takes about two minutes longer than typing "summarise this", and the difference in what you get back is not subtle.