How to summarize a long document without losing the point

Will it fit?

Models measure text in tokens, not pages. OpenAI’s rule of thumb for English is that a token is about three quarters of a word. With that, the arithmetic for a typical work document is short.

The current top models from OpenAI, Anthropic and Google accept around a million tokens through their APIs, according to the thinkhard.si model catalog checked October 20, 2026. Chat apps can allow less, depending on your plan, and may not tell you when a file was cut short. Two quick ways to know where you stand: count the text in the Token Counter before you paste, and after you upload a file, ask for the heading of the last page. If the model cannot give it, it did not read that far.

A scanned PDF is a picture of text, not text. Most assistants can read page images now, but numbers and tables come through less reliably, so check figures from scans with extra care.

Why long summaries miss things

Fitting is not the same as reading evenly. In a widely cited study, “Lost in the Middle”, researchers found that language models used information best when it sat at the beginning or end of a long input, and worse when it sat in the middle, even models built for long inputs. Models have improved since that study and the size of the effect varies by model, but the vendors’ own advice for long documents still follows from it.

Two other things cost a summary its accuracy. A summary with no stated reader keeps whatever the model finds typical of the genre, so a security review gets summarized like any security review, not for the decision you face. And compression drops the small words that matter most: “unless”, “only if”, “up to”, the exception in a footnote. Each step below closes one of these gaps.

1. Say who it is for, and what it feeds

The same 48-page policy needs a different summary for the finance team, for team leads who approve expenses, and for a new hire. Name the reader, what they already know, and what they will do with the summary. That one line decides what the model keeps and what it drops, which is the whole job of a summary. If you need versions for several readers, the three audiences template writes them side by side from one reading.

2. Document first, request last

Put the document at the top of the prompt, then what you want. Anthropic’s prompting guidance says that with long inputs, a query placed after the documents can improve response quality by up to 30% in its tests, and Google’s long-context guide gives the same advice for Gemini. When you upload a file instead of pasting, the file comes first anyway; type your request after attaching it.

3. Quotes first, then the summary

Asking for quotes before the summary does two jobs. It makes the model go back through the document for the passages that matter instead of summarizing from its overall impression, which is the habit Anthropic’s guidance recommends for long documents. And it gives you a page number for every point, so checking takes minutes rather than a full reread.

Summarize this report.

{{report}}

"""
{{policy}}
"""

Above is the 48-page draft of our new travel and expense policy. It replaces the 2024 policy, pasted below it.

"""
{{old_policy}}
"""

Readers: our 12 team leads, who must apply the new policy from January 1 and will read this on their phones.

First, find the passages in the new policy that change what a team lead approves or checks, and quote each one with its page number.
Then, using only those quotes, write:
- One line on the biggest change.
- Up to six changes as bullets, each with its page number.
- What a team lead must do differently from January 1.
Keep it under 300 words. Leave out changes that only affect the finance team.
Example of a bullet: "Hotel limit rises from $180 to $220 a night in high-cost cities (p. 14)."
A good summary lets a team lead approve a claim correctly without opening the policy.
If the policy does not say something a team lead needs, write "Not in the policy" instead of guessing.

What changed. The first prompt gets an even-handed tour of the whole document. The second gets the six things twelve specific people must do differently, each with a page you can check, and an honest “Not in the policy” where the draft is silent, which is often the most useful line for whoever is still editing it. The hotel figures are an invented example of format; the model fills in the real ones.

For the most common versions of this job, ready prompts are in the research and long documents collection: the executive summary template for a decision-maker with two minutes, and the questions from a long document template when you need specific answers, each with a page, a short quote and a plain “not found” where the document is silent.

4. Check the quotes, then the gaps

Checking a quote-based summary is fast because it points at its own evidence. Pick the three points that matter most, open the pages they cite, and confirm the quote is there and says what the summary claims. Watch especially for numbers, dates and words like “only”, “unless” and “up to”, which compression tends to lose. The Fact Check List pulls every number, date and name out of a summary so none slips past.

Then ask about what is not there. Two follow-ups in the same chat catch most omissions:

What did you leave out that a team lead would want to know? List it with page numbers.

Does anything in the middle sections (pages 15 to 35) change or contradict the summary?

The second question targets the middle of the document on purpose. The guide to reducing hallucinations explains why “Not in the document” as an allowed answer matters as much as any instruction to be accurate.

Very long documents, or many

For a document too long for your plan, or a pile of reports at once, split the work. Ask first for a map: the table of contents and one line on what each section covers. Then summarize each section with the same brief, so every part is treated the same way, and finally ask for one summary built from the section summaries, for your reader. Each step stays short enough to read evenly. The articles to digest template does the many-documents version, and the guide to extracting data covers documents where you need exact fields rather than a summary.

If the document is confidential, such as a contract, a personnel file or a client’s figures, use the AI tool your employer approves for it. For pasted text, Safe Paste swaps names and numbers for placeholders first.

A checklist to copy

Before you prompt
[ ] Who reads the summary, and what they will do with it
[ ] Does it fit? Count it, or ask for the last page's heading
[ ] Confidential? Use the approved tool, or anonymize

In the prompt
[ ] Document at the top, request at the end
[ ] Quote the passages that matter, with page numbers
[ ] Summary built only from those quotes
[ ] Length and shape for the reader
[ ] What to leave out
[ ] "Write 'Not in the document' instead of guessing"

After the answer
[ ] Three most important points checked against their pages
[ ] Numbers, dates, "only", "unless" and "up to" checked
[ ] "What did you leave out?"
[ ] "Does the middle change anything?"

Related

• Answers from a long documentEach answer with a page, a short quote and a plain “not found”.Template
• Token CounterCheck a long document fits before you paste it.Free tool
• How to fact-check an AI answerCheck the numbers, dates and quotes in any summary.Guide
• AI meeting minutesThe same habits for transcripts: decisions, owners and dates.Guide