Summarising a Long Report Without Losing the Point

A one-shot summary reliably keeps the topic and drops the point. The fix is to make extraction a separate step from writing, so the figures, the dates and the hedges are captured before anything starts compressing them.

Why asking for a summary fails

Ask for a summary of a hundred-page report and you get something that reads well and is nearly useless: it tells you the report is about supply chain resilience, that several risks were identified and that the authors recommend further monitoring. Every specific has been smoothed away.

This is not a defect to be prompted around; it is what summarising means when nobody has said what matters. Compression must discard something, and with no instruction the discarded material is whatever is least common in the text — which is precisely the figures, the exceptions and the conditions, because those appear once each while the topic appears on every page.

The second reason is structural. If the tool is using retrieval rather than reading the whole file, a “summary” is a summary of whichever passages happened to match, presented as a summary of the document. Establish which you have before you start — the two tests are in getting an AI to work with your own documents.

Pass one: extract, do not summarise

The first pass produces no prose. It produces a structured list of things found in the document, each tied to a location. Work section by section for anything over about thirty pages.

Read section 3 of the attached report. Do not summarise it.

Extract, as a list, using only what is in the text:

CLAIMS: each substantive claim, one line each, with the sentence it
  comes from quoted and its page number.
FIGURES: every number that carries meaning - amounts, percentages,
  dates, counts. Give the number, what it measures, the period it
  covers, and the page.
CONDITIONS: every hedge, assumption, exclusion or limitation attached
  to a claim or figure ("excluding", "assuming", "in the pilot region
  only", "provisional", "before tax").
DISAGREEMENTS: anywhere the text reports that people disagree, or
  where a figure is given more than one value.
UNKNOWNS: anything the text says is not known, not measured or still
  under review.

If a category is empty for this section, write "none". Do not
infer, do not round, do not tidy the wording of a quote.

The output is ugly and long. That is correct — it is working material, not a deliverable. Its value is that everything in it is traceable to a page, so the second pass has nowhere to invent from.

Do this for each section and paste the extracts together. For a hundred-page report this is perhaps six passes and twenty minutes, which is still a fraction of reading it.

Pass two: compose from the extract only

Start a fresh conversation. Paste the combined extract and not the original document. This is the load-bearing detail: if the source is still available, the model will drift back to it and the extraction discipline is lost.

Below is a structured extract from a report. Write a summary for
a reader who will use it to decide whether to approve a budget.

Rules:
- Use only what is in the extract. Nothing else exists.
- Every figure you state must carry the qualification attached to it
  in the extract. If a figure was "provisional" or "excluding
  region 4", those words come with it.
- Where the extract records a disagreement or an unknown, the
  summary says so explicitly. These are the most decision-relevant
  lines in it.
- 500 words maximum. Plain paragraphs. No bullet points except in
  the figures block.
- End with a section headed "What this report does not tell you",
  drawn from the UNKNOWNS entries.

[paste extract]

The closing section is the part readers use. A summary that ends with four honest sentences about what was not measured is worth more to somebody deciding than three extra paragraphs of findings, and no one-shot summary has ever produced one.

Keeping the caveats

Qualifications are attached to figures in the original and detached from them in every careless summary. That detachment is how a “projected 30% saving, pre-implementation-cost, in one region, under the optimistic scenario” becomes “30% saving” in a board pack and a problem eighteen months later.

The rule that prevents it is mechanical: a number and its qualifier are one unit and never travel separately. Enforce it by writing figures in a fixed form during extraction.

30% | cost reduction | Northern region pilot only | projected,
     not observed | excludes implementation cost | p.47

18 months | payback period | assumes 2026 energy prices hold |
     stated as "roughly" in the text | p.51

A figure that cannot be written in that form — because you cannot find what it measures or what period it covers — is a figure not to quote.

The three-minute check

The method reduces errors; it does not eliminate them. Three checks catch nearly all of what survives.

1. Search for three quoted sentences in the original. Pick the three that matter most. A quote that is not in the document invalidates the whole extract, not just that line, and you restart the pass.
2. Check the largest number and the smallest. Extremes are where transcription errors and misplaced decimal points land, and they are also the ones that get repeated.
3. Read the document’s own conclusion and compare. If the report’s final section says something your summary does not, find out why. This catches the failure the method is most prone to: an extract that is accurate section by section and misses the argument that spans them.

Three variants worth having

• The decision brief. Compose against the question being decided rather than the document: “organise this around whether to renew the contract; anything that does not bear on that goes in an appendix list of headings only”.
• The change summary. For a new version of a document you already know, extract from both and ask only for differences. Far more useful than a fresh summary, and much shorter.
• The objection list. “From the extract, list the five weakest points in the argument and what evidence would be needed to shore each one up.” Useful before you have to respond to the document in a room. See preparing for a meeting.

Transcripts are a different problem

A meeting or interview transcript looks like a document and behaves like a different thing. Reports are written to be summarised: they have headings, a structure and a conclusion. Transcripts have none of that. The important sentence is often eleven words long, in the middle, buried in an exchange about something else, and the extraction schema above will walk past it.

Two changes make the method work. Extract by speaker act rather than by claim, and extract the disagreements rather than the topics.

From this transcript, extract - do not summarise:

DECISIONS: anything settled. Quote the exact words that settled it
  and who said them.
COMMITMENTS: anyone saying they will do something. Who, what, by
  when. Quote it. Mark [NO DEADLINE] where none was given.
DISAGREEMENTS: where two people wanted different things. Both
  positions, in their words, and whether it was resolved.
UNANSWERED: questions asked and not answered. These are the most
  useful lines in any transcript and they are always dropped.
CONDITIONS: anything agreed subject to something else.

Ignore pleasantries, repetition and anything not in these five
categories. Do not infer intent from tone.

The unanswered-questions category earns its place. In a recorded meeting, a question that is deflected and never returned to is usually the substantive one, and no ordinary summary contains it because nothing was said about it worth summarising.

One warning specific to this material: automatic transcripts misattribute speech, especially where people talk over each other, and a quotation lifted from one may be assigned to the wrong person. Check any quote against the audio before repeating it as what somebody said.

When not to summarise at all

Do not summarise a document whose exact wording is the point. Contracts, statutes, regulatory correspondence, medical instructions, safety procedures: for these the correct operation is to locate the relevant clause and read it. A summary of a contract is a paraphrase of an obligation, and paraphrases of obligations are not obligations.

Do not summarise something you will be answerable for having read, unless you also read it. This is not a moral point but a practical one: the questions people ask you afterwards are about the details a summary removed.

And do not summarise a document that is already a summary. Executive summaries and abstracts have been compressed once by somebody who knew the material; compressing again mostly removes the remaining qualifications.