CatchCase

Summarise a long PDF without missing details

"Summarise this PDF" is one of the most common things people ask an AI assistant, and one of the easiest to get subtly wrong. The summary reads well. It's confident and organised. And it leaves out the clause on page 34 that changes everything, or rounds a figure from the wrong table.

The problem isn't that AI can't summarise. It's that "summarise" is an open instruction, and a long document gives a model plenty of room to choose what to keep. Here's how to narrow that room so the summary is one you can rely on.

1. Make sure the file was actually read

Before anything else, check the model can see the whole document.

  • Scanned PDFs (photos of pages) often have no text layer. Try selecting text in your PDF viewer; if you can't, the file needs OCR (text recognition) first, or the model may only see a few fragments.
  • Very long files can exceed what a tool reads in one go. Some tools silently use only part of the document.
  • Tables and charts may come through as jumbled text.

A quick test: ask for a map first.

Before summarising: list the sections of this document with their page ranges, and any tables or appendices.

If the outline stops at page 40 of 120, or shows a section you know isn't there, you've caught the problem before it reached your summary.

2. Tell it who the summary is for

A summary for a CFO, a project team and a student are three different documents. Say who will read it and what they need to decide:

I'm the operations manager. I need to know what this audit means for my team: findings that affect operations, deadlines, and anything we must change.

That single sentence does more for relevance than any formatting instruction.

3. Ask for structure, not prose

Paragraph summaries hide omissions. A fixed structure exposes them, because an empty section is visible:

  • TL;DR — three sentences: what the document is, what it concludes, what action is needed.
  • Key findings — no more than eight bullet points.
  • Decisions made — if any.
  • Open questions and risks — what's unresolved.
  • Action items — with owners and deadlines, if the document gives them.

Our PDF report summary case uses exactly this structure, so you can run it on any long report without writing the prompt yourself.

4. Require page references

This is the single most effective habit. Ask for a page number (or section number) after every point:

After each finding, give the page number where it appears. If a point combines information from several pages, list them all.

Two things happen. The model is pushed to anchor every claim in the text, which reduces invented details. And checking becomes fast: you jump to the page instead of searching the whole file.

5. Split long documents by section

For documents over about 50 pages, a single summary has to compress too much. Work in stages:

  1. Get the section map (step 1).
  2. Summarise each major section separately, with page references.
  3. Ask for an overall summary built from the section summaries.

It takes a few more messages, but each step is small enough to check, and nothing gets squeezed out in the compression.

6. Ask the questions a summary won't answer

Summaries favour what the document emphasises. The details that matter to you are often in the fine print. After the summary, ask narrow questions:

  • "List every date or deadline in this document and what happens on it."
  • "Which parts mention penalties, fees or liability? Quote them."
  • "What does the document say about [your topic]? If it says nothing, say so."

"Not in the document" is a useful answer. It's often exactly what you needed to know.

For contracts specifically, a general summary isn't enough — you want the clauses that can hurt you. The contract risk review case reads a contract from your side of the deal and lists risky clauses by clause number, in plain language, with wording to ask for instead.

7. Verify the numbers that matter

Figures are where summaries most often go wrong: a number from the wrong year's column, a percentage that's actually a percentage-point change, a total that combines two tables. For any figure you'll repeat to someone else:

  1. Go to the page reference.
  2. Check the number, the unit and the period.
  3. Check it's the figure you think it is (actual vs forecast, gross vs net).

It takes a minute per figure, and it's the minute that protects your credibility.

A worked example

Say you've been sent a 90-page annual report from a supplier and need a briefing for your manager. A good sequence:

  1. Map: "List the sections with page ranges." → 9 sections, financials on pp. 52–80.
  2. Focused summary: "I manage this supplier relationship. Summarise what affects us: delivery performance, pricing, risks to supply. Page reference after each point."
  3. Narrow questions: "Does the report mention any changes to pricing or contract terms for 2027? Quote them." → "Page 14: 'We expect to review freight surcharges in the first half of 2027.'"
  4. Verify: check the on-time delivery figure on page 23 is for your region, not the whole company.
  5. Rewrite for your audience: "Turn this into five bullet points for my manager."

The whole thing takes fifteen minutes, and every claim in the final briefing can be traced to a page.

For academic papers

Research papers need a different structure: research question, method, findings with the actual numbers, limitations, and a correct reference. The emphasis is on not overstating results. The research paper summary case asks for exactly that, keeps the authors' level of certainty, separates the limitations the authors state from the ones a careful reader would add, and formats the reference without inventing a missing DOI.

Common mistakes

  • Trusting the first summary of a scanned file. Check the map.
  • "Summarise this" with no context. You get the average summary, not yours.
  • No page references. You can't check what you can't find.
  • Forwarding numbers unchecked. Verify every figure you'll repeat.
  • Asking one summary to do everything. Split long documents and ask narrow follow-ups.

The short version

Check the file was read, say who the summary is for, ask for a fixed structure with page references, split long documents, ask narrow questions about what matters to you, and verify the key numbers. With those habits, a summary stops being a guess about what's important and becomes a reliable map of the document.