Prompt Engineering

Document Summary Prompts: 9 Templates to Summarize Long Documents with AI

To summarize a long document with ChatGPT or Claude, a good prompt states who will read the summary, what decision it supports and what form it should take. It also asks the tool to cite the passages it relies on.

RRomain BellaicheNovember 19, 202616 min read

Read the full guide

Prompt Engineering 2026: Complete SME Guide (ChatGPT, Claude, Gemini, Copilot)

Document Summary Prompts: 9 Templates to Summarize Long Documents with AI

To summarize a long document with ChatGPT or Claude, a good prompt states who will read the summary, what decision it supports and what form it should take. It also asks the tool to cite the passages it relies on. Without those instructions, you get a summary that reads well, looks balanced, and is often incomplete exactly where it matters.

This article is for SME owners, HR managers and admin teams who receive reports, tenders, contracts, minutes or meeting transcripts every week that are too long to read in full. You will find how to prepare a document before handing it to AI, nine ready-to-use prompts, a method for texts that exceed what the tool can handle, and a review routine. The basics of writing instructions are covered in the prompt engineering guide for SMEs, the starting point of this series.

Why AI summaries of long documents go wrong

A language model does not read an 80-page document the way you do: it processes text split into tokens, gives more weight to the beginning and the end, and produces something that looks like a summary even when it has missed the key point. Knowing these three mechanisms helps you write prompts that compensate for them.

First, the size limit. Every LLM handles a maximum amount of text at once, measured in tokens. That limit varies by tool and plan. When a document goes beyond it, some tools truncate without warning or read only part of the file. You then get a summary of the first 30 pages presented as a summary of the whole thing.

Second, uneven attention. Stanford researchers showed in a 2023 study titled "Lost in the Middle" that models retrieve information placed at the beginning or end of a long text better than information in the middle. Recent models have improved on this, but the habit still makes sense: a termination clause on page 47 is more likely to be missed than a figure in the introduction.

Third, smoothing. A generic summary follows the document's structure and gives each part space in proportion to its length. Yet what matters to you sometimes fits in two lines: a penalty, a deadline, a condition precedent. And when the tool fills a gap with a plausible sentence, you are in the territory of AI hallucinations, which are all the harder to spot because you have not read the original.

Preparing the document before you paste it

The quality of a summary depends as much on the file you provide as on the prompt: a clean PDF, split along its structure and stripped of unnecessary data, gives much better results than a raw scan. Five minutes of preparation is enough in most cases.

What to check:

  • Can the machine read the text? A scanned PDF is an image. If you cannot select the text in your PDF reader, run optical character recognition (OCR) first, otherwise the tool is guessing.
  • Do tables survive copy and paste? Financial tables and price grids often get scrambled. For those parts, send the source file (Excel, CSV) or a screenshot instead, and use the methods in our data analysis prompts.
  • Does the document have a structure? Find the table of contents, part headings and appendices. You will need them to split a long text and to ask for precise references.
  • Is all of it needed? Standard appendices, cover pages and repeated notices take up space for nothing. Remove them if they do not serve your question.
  • Is there personal data? Employee names, customer contact details, salaries: replace them with placeholders ("Employee A", "Customer 1") when the summary does not need them.

If you often summarize the same type of document (board minutes, public tenders, monthly reports), save your framing as a standing instruction in a Claude project, a custom GPT or the equivalent in your tool. What belongs in standing instructions versus the one-off request is explained in our article on system prompts and user prompts.

The 9 prompts for summarizing long documents

Each template follows the same structure: context, reader, purpose of the summary, format constraints and a citation requirement. Replace the items in square brackets. The logic behind this structure is explained in the anatomy of a good prompt.

1. Summary framed for the reader

I am [role] at a [type of company, headcount]. Here is a [number]-page document: [attach or paste]. Summarize it for [reader: me, my business partner, the leadership team] who needs to [purpose: decide on, prepare, respond to]. Length: [number] words maximum. Start with the main conclusion in two sentences. For each key point, give the page or section it comes from in brackets. If any part of the document could not be read or seems incomplete, say so before the summary.

The last line is the most useful one. It forces the tool to flag truncation instead of hiding it.

2. One-page decision memo

Here is a [report, audit, commercial proposal]: [attach]. Write a decision memo of one page maximum, in four parts: what is being proposed; key figures (amounts, deadlines, volumes), copied exactly as they appear in the document; risks or conditions mentioned; open questions. Do not add any recommendation that is not in the document. If you make one, put it in a separate section titled "AI opinion, to be checked".

Keeping what the document says apart from what the tool thinks stops you from presenting a generated opinion to your board as the auditor's conclusion.

3. Structured extraction into a table

Read the attached document and extract into a table every [date, deadline, amount, obligation, commitment] it contains. Columns: item, exact value, who is concerned, page or section, short quote from the passage. Do not rephrase values: copy them. End with the total number of items found.

This is often more useful than a prose summary: a 60-page tender shrinks to around twenty rows of dates, documents to supply and award criteria. The total acts as a check: if you rerun the prompt and get a different number, something was missed.

4. Section-by-section summary, then synthesis

Step 1: here is part [number] of [total] of the document [title]: [paste the part]. Summarize it in [number] points, with figures and page references. Do not write an overall synthesis yet. [Repeat for each part, in the same conversation or by collecting the summaries.] Final step: here are the summaries of the [total] parts: [paste]. Write a [number]-word synthesis that brings out common points, contradictions between parts and the three most important items for [purpose].

This is the most reliable method for a document that exceeds what the tool can handle, or whose middle matters as much as its beginning. It takes longer, but every part gets the same attention.

5. Reading a contract or terms and conditions

Here is a [contract, terms and conditions, lease] my company is about to sign as [customer, supplier, tenant]: [attach]. List, with the clause reference and a short quote for each: term and renewal conditions; termination conditions and notice periods; penalties and liability caps; confidentiality, intellectual property and non-compete clauses; any commitment that is unusual for this type of contract. This is a preparatory reading, not legal advice: flag the clauses that should be reviewed by a lawyer.

This prompt saves you time before a meeting with your lawyer or accountant. It does not replace them. For documents in a foreign language, combine it with our translation prompts, especially the working translation template.

6. Minutes from a meeting transcript

Here is the transcript of a [duration] meeting between [participants and roles]: [paste]. Write minutes covering: decisions made; actions, each with an owner and a deadline if mentioned; open points; disagreements expressed. Only assign a decision or action to a person if the transcript states it clearly. Otherwise, write "owner not specified".

Automatic transcripts often mix up speakers and proper names. The final instruction stops the tool from inventing who committed to what, which is the most awkward error once minutes have been sent out.

7. Comparing two versions or several documents

Here are two versions of the same document: version A [attach] and version B [attach]. List every substantive difference (amounts, dates, obligations, scope) in a table: section, version A, version B, impact of the change. Ignore formatting and style differences. If you cannot confirm that a section is identical in both versions, flag it as "to be checked".

The same template works for comparing three quotes or two audit reports: replace "versions" with "documents" and spell out the comparison criteria.

8. Questions and answers grounded in the text

I am going to ask you questions about the attached document. Rules for all your answers: answer only from the document; quote the exact passage that supports your answer, with the page or section; if the document does not contain the answer, write "not found in the document" and do not use your general knowledge. First question: [question].

Rather than asking for an overall summary, you question the document about what you actually need. It is often faster and more reliable, especially if you add examples of the answers you expect, following the few-shot prompting principle.

9. Faithfulness check of a summary

Here is a source document [attach] and a summary of it [paste]. Check the summary sentence by sentence. For each sentence, state: "confirmed" with a quote from the source passage, "not found" if nothing in the document supports it, or "distorted" if the document says something else, with the quote. Then list the important information in the document that is missing from the summary.

Run this check in a new conversation, ideally with a different tool from the one that wrote the summary. Asking for a sentence-by-sentence review, as in chain of thought prompting, makes discrepancies easier to see.

Very long documents: split, attach or move to RAG

Up to a few dozen pages, an attached file and a well-framed prompt are usually enough; beyond that, or when you query a set of documents regularly, splitting and then RAG become more reliable. The right choice depends on volume, frequency and how much accuracy you need.

SituationRecommended methodWatch out for
Document under 30 pages, one-off useAttached file + prompt 1, 2 or 3Check that the whole file was read
Long document where every part matters (report, tender)Split by section + prompt 4Use the same instruction for each part
Several documents to cross-checkProject or file space + prompt 7 or 8Name files clearly
Document base queried every week (procedures, contracts, archives)RAG: search the documents, then answerTechnical setup, to be scoped with a specialist

The 30-page threshold is a practical guide, not a technical limit: it depends on how dense the text is and on the tool. What matters is checking, every time, that the tool has processed the whole document.

The difference between pasting a document into a conversation and building a document search system is explained in our article RAG vs prompting. To choose an assistant based on accepted file size and project features, see our ChatGPT vs Claude vs Gemini comparison, along with our pages on using Claude in SMEs and ChatGPT in SMEs. If your company works in Microsoft 365, our Copilot 365 beginner's guide covers summarizing Word documents and Teams meetings.

Checking a summary before you rely on it

The level of review should match what is at stake: a summary for information can be skimmed, a summary used to sign, hire or commit spending must be checked against the original. A three-level grid covers most SMEs.

  1. Staying informed (market watch, article, industry report): prompt 1 with page references, and read the cited passages for the two or three points you keep.
  2. Preparing a decision (commercial proposal, audit, tender): prompt 2 or 3, then a faithfulness check (prompt 9) in a new conversation.
  3. Committing the company (contract, HR document, reply to a public authority): the summary prepares the reading, it does not replace it. Flagged clauses are read in the original, and a professional approves where needed.

Three checks take less than five minutes at any level: every figure in the summary appears identically in the document; dates and deadlines have not been rounded or shifted; no qualifier has disappeared ("subject to", "up to", "unless otherwise agreed").

A document summarized by AI is still a company document: contracts, employee files and customer data should only go through a properly configured business account, in line with your AI policy. Summarizing is the use where people most readily upload entire files without thinking.

The GDPR (Regulation (EU) 2016/679) applies as soon as a document contains personal data, and its data minimization principle requires processing only the data that is necessary. In practice, an annual review report or a job application file should not be pasted into a consumer tool without anonymization. Also check in your plan settings whether your conversations can be used to train models. The habits to adopt are detailed in our article on data confidentiality with AI, and the risks of personal accounts used at work in our article on shadow AI.

On the EU side, Article 4 of the AI Act (Regulation (EU) 2024/1689), as amended by Regulation (EU) 2026/1744 of 8 July 2026, requires companies deploying AI systems to take measures to support AI literacy among their staff, taking into account their knowledge and the context of use, without imposing a specific level of competence. Teaching teams to check a summary before relying on it is part of those measures. Details for using ChatGPT at work are in our article on AI Act compliance for ChatGPT.

Frequently asked questions

What is the best prompt for summarizing a document with ChatGPT?

There is no single formula. A good summary prompt specifies the reader, the purpose, the length, the expected format, and requires page references. Template 1 in this article is a reusable starting point for most documents.

Claude or ChatGPT: which one summarizes a long PDF better?

Both can summarize long documents; the differences lie mainly in accepted file size, project features and privacy options depending on the plan. Test them on one of your real documents with the same prompt, then compare using prompt 9.

How many pages can you summarize at once?

It depends on the tool, the plan and how dense the text is. Beyond a few dozen pages, split the document by section (prompt 4) and always check that the tool confirms it has read the whole thing.

How do I know if the AI made something up in the summary?

Require quotes and page references from the first prompt, then run a sentence-by-sentence faithfulness check in a new conversation (prompt 9). Any sentence without an identifiable source passage should be treated as unreliable.

Can I summarize a contract with AI before signing it?

Yes, to prepare your reading and spot sensitive clauses (prompt 5). No, as a substitute for reading the key clauses or getting legal advice: the company is bound by the signed text, not by its summary.

Can I summarize documents that contain personal data?

Only with a properly configured business account, applying the GDPR data minimization principle: remove or replace any data the summary does not need. For HR files, anonymization should be systematic.

How GrowthPerf can help

Summarizing with AI saves hours, provided you have a method: prepare the document, frame the prompt, require sources and check according to what is at stake. GrowthPerf, a Qualiopi-certified training provider, trains leaders and teams at SMEs, startups and non-profits in these practices, using their own documents.

The Operational AI course includes workshops on your real files: a library of summary and extraction prompts, a review grid, and confidentiality rules suited to your business. For a programme focused on prompting techniques, see our page on prompt engineering training. Depending on your situation, these courses can be funded by your OPCO (French training fund).

For the full set of techniques, go back to the complete prompt engineering guide for SMEs, and to the glossary for the basics.

Are your teams spending too much time on reports, tenders or minutes? Book a free 30-minute audit: we test these prompts together on one of your documents and define the right level of review.

R
Romain Bellaiche

AI & No-Code Expert

Assess your AI maturity in 3 minutes

Answer a few questions and get a personalized assessment with recommendations tailored to your industry.

Assess my AI maturity