How to summarise long documents with AI without losing what matters

Summarising is probably the most common use of AI at work, and also one of the most treacherous. A summary always sounds good: clear sentences, tidy points, a confident tone. The problem is what does not appear. A penalty clause that was left out, a figure that got rounded or a conclusion in the appendix that contradicts the executive summary.

What follows is a method that works with any assistant and any kind of long document: reports, contracts, regulations, transcripts, theses or manuals.

Why long documents are hard

Two technical reasons explain why an 80-page document is not summarised the same way as a five-page one.

• Attention is not spread evenly. Models process the whole text at once, but they do not make equal use of every part. We explain this in Tokens and the context window: information in the middle of a very long text tends to get less attention than what is at the start or end.
• Compressing forces choices. Reducing 40,000 words to 500 means discarding 99%. If you do not tell the model what matters to you, it will choose by its own criteria, which tend to favour the general over the specific.

That is why the method rests on two ideas: telling it what to look for, and being able to check where each point comes from.

The five-step method

1. Define what the summary is for

Before pasting anything, decide what you will do with the result. These are not the same:

• Deciding. You need options, costs, risks and a recommendation.
• Briefing other people. You need the main facts in plain language, without jargon.
• Reviewing obligations or risks. You need deadlines, commitments, penalties and conditions, without leaving any out.
• Studying or learning. You need the key concepts, how they relate and examples.

Tell the model in one sentence: "I am going to use this summary to decide whether we renew the contract with this supplier."

2. Put the document first and the request last

With long texts, place the document at the start of the message and your instructions at the end. Anthropic's documentation on long contexts recommends this explicitly, and states that in its tests putting the question at the end improved response quality. If you upload a file, write your instructions after attaching it.

If you paste several documents, separate and label them clearly ("Document 1: framework contract", "Document 2: pricing annex") so the model can refer to each one.

3. Ask for a specific structure

A defined format avoids the generic summary and makes gaps visible. For example, for a contract:

• Parties and purpose of the contract, in two lines.
• Deadlines and key dates, with the clause they come from.
• Each party's obligations, separately.
• Payments, penalties and exit conditions.
• Ambiguous points or ones that need review by a professional.

For a report, a useful structure is: main conclusion, three to five findings with their figure, limitations the report itself mentions and what it does not answer.

4. Require page or section references

This is the instruction that most changes how useful the result is. Ask for each point to end with the page, section or clause it comes from, for example "(p. 14)" or "(clause 7.2)". That way you can check any point in seconds, and the model tends to be more faithful when it knows it has to point to the source.

If something important is not in the document, ask it to say so explicitly instead of filling it in: "If the document does not mention the notice period, write 'Not stated'."

5. Check what you are going to use

You do not need to reread the whole document. Check against the original:

• Every figure and date you are going to quote or use to decide.
• The names of people, companies and products.
• Obligations and conditions, word for word if they have legal or financial consequences.
• One or two random claims, to get a sense of overall accuracy.

For a fuller method, see How to check an AI answer before you use it.

When the document is very long: summarise in parts

If the document is more than your tool accepts, or if you notice the summary getting vague, split it. The technique is known as chunked summarisation (chunking) and works like this:

• Split by natural sections, such as chapters or headings, not blindly by page count.
• Summarise each section with the same structure and the same instructions.
• Combine the partial summaries in a final step and ask for the overall summary based on them, keeping the references.

It is slower, but it greatly reduces omissions, because at each step the model works with a text it can attend to fully.

A prompt to get started

[ATTACH OR PASTE THE DOCUMENT HERE]

Context: I am going to use this summary to [DECIDE / BRIEF / REVIEW / STUDY].
Who will read it: [AUDIENCE].

Summarise the document above with this structure:
1. Main conclusion (3 sentences maximum).
2. Key points (between 5 and 8), each with its specific figure or fact.
3. Risks, obligations or warnings it mentions.
4. What the document does not answer or leaves ambiguous.

Rules:
- End each point with the source page or section in brackets.
- Do not add information that is not in the document.
- If a detail I ask for does not appear, write "Not stated".
- Copy figures, dates and names exactly.

There is a version adapted to technical documents in the Technical PDF summary template.

Common mistakes

• Asking for "a summary" and nothing more. You will get a generic one that treats the important and the incidental the same way.
• Relying on the document's own executive summary. The model may lean on it too heavily. If you care about the appendices or the small print, ask for them explicitly.
• Summarising in a long conversation. If you already have many messages, start a new conversation for the document: earlier instructions may interfere.
• Assuming it read everything. Some tools truncate very long files without warning. Ask which is the last section it has available and compare it with the original.

Keep reading

• Tokens and the context window: why long texts are processed less well.
• Why AI makes things up: how to reduce invention when working with documents.
• How to check an AI answer before you use it: what to review before sharing a summary.
• What never to paste into an AI chatbot: what to do with confidential documents.
• In the glossary: context window, chunking and OCR.

Sources

• Anthropic. Long context prompting tips. Claude documentation.
• Google. Long context. Gemini API documentation.
• Liu, N. F. et al. (2023). Lost in the Middle: How Language Models Use Long Contexts. arXiv.