How to summarize documents with AI: PDFs, transcripts, reports, and email threads

Why "summarize this" is the wrong prompt

"Summarize this document" produces a smooth paragraph that reads well and tells you almost nothing you can act on. Worse, it hides what was left out. A prose summary of a 40-page report has no gaps you can see, because gaps in prose are invisible.

The fix is to stop asking for summaries and start asking for extractions: the decisions, the actions with owners and dates, the numbers, the open questions, the things that changed since last time. When those come back as labeled lists, a missing item is a blank line you can notice.

This guide gives you the method: the structured prompt, how to handle documents too long to paste, transcripts and email threads, the omission check that should follow every summary, and the tools that ground their answers in your files so you can click through and verify.

Start with the question, not the document

Before you paste anything, decide what you need from it. "I need to know whether this contract lets us terminate early and what it costs" is a question. "Summarize this contract" is a wish. The first produces a two-paragraph answer with clause numbers; the second produces a page about the parties and the effective date.

Three questions that cover most work documents:

• What was decided, and what is still open?
• What do I have to do, by when, and what do others owe me?
• What changed compared with the last version, meeting, or quarter?

Put the question at the top of the prompt, before the document, and again at the bottom. Long pastes push instructions out of the model's attention; repeating the question after the document fixes most of that.

The structured summary prompt

This is the workhorse. It works on reports, transcripts, proposals, and threads, and it produces output you can check.

You are a meticulous analyst. I will give you a document. Do not write a prose summary. Instead, extract the following, using only what is in the document. If a section has nothing, write "None found." Never infer or fill gaps.

1. Purpose: one sentence on what this document is and who wrote it for whom.
2. Decisions made: each as one line, with who decided (if stated) and the page or section where it appears.
3. Action items: a table with columns Owner, Action, Due date, Source location. Use "not stated" where the document does not say.
4. Numbers and dates: every figure, amount, deadline, and date that appears, with its context and location.
5. Open questions and unresolved issues: things the document raises but does not settle.
6. Risks, caveats, and conditions: anything phrased as "if," "unless," "subject to," or "pending."
7. What I should read in full: the two or three sections that matter most for my question, with a one-line reason each.

My question: [WHAT I NEED TO KNOW OR DECIDE]

Document:
[PASTE THE DOCUMENT OR ATTACH THE FILE]

Reminder: extract only. Cite the page or section for every item.

The Source location column and the page citations matter. They turn the summary into an index. When something looks important or surprising, you go to the page and read it yourself.

Long documents: chunking and running summaries

How much fits

Every model has a context window: the amount of text it can hold at once, measured in tokens (a token is roughly three quarters of an English word). The current major assistants handle long documents, often hundreds of pages, but two things go wrong before the hard limit. Quality drops on the middle of very long inputs, and a tool may silently truncate a file that is over its limit and summarize only the first part.

Test this. Ask "what is the last section heading in the document I gave you?" If the answer is a heading from page 30 of a 120-page file, you have been summarizing a fragment.

The chunk-and-merge method

For anything over roughly 50 pages, or anything where completeness matters (contracts, audit reports, due diligence), chunk it:

1. Split the document at natural boundaries: chapters, numbered sections, meeting dates. Aim for pieces you would be comfortable reading in one sitting.
2. Run the structured prompt above on each piece, in a fresh chat or with a clear "Part 3 of 7" label, so the model does not blend parts.
3. Paste the structured outputs (not the source) into one final prompt: "Below are extractions from seven parts of one document. Merge them into a single set of decisions, actions, numbers, and open questions. Flag any item that appears in two parts with different details."
4. Spot-check three items against the source pages. If any of the three is wrong, redo the chunk it came from with a smaller piece.

It takes longer than one paste. It is also the difference between a summary you can send to your boss and one you cannot.

Scanned PDFs

If the PDF is a scan, the model may see it as images. Most assistants now run text recognition (OCR) on uploads, but results degrade on poor scans, handwriting, and tables. Ask "was this document readable as text, or did you interpret images?" and, for anything numeric, verify against the page.

Meeting transcripts

Transcripts from Otter.ai, Fireflies, or Zoom AI Companion come with built-in summaries. They are decent for "what did we talk about" and unreliable for "what exactly did we agree." Speaker labels get swapped, and a tentative "we could probably do that by Friday" becomes "Action: deliver by Friday."

Run the transcript through the structured prompt with two additions:

This is a meeting transcript. Speaker labels may be wrong; when a statement's owner is unclear, write "unclear (possibly [NAME])" rather than assigning it. Distinguish firm commitments ("I will send it Tuesday") from tentative ones ("we could maybe") and label each action FIRM or TENTATIVE. Quote the exact words for every action item so I can verify.

Transcript:
[PASTE TRANSCRIPT]

Then send the action list to the attendees the same day, while memory is fresh enough to correct it. That step catches more errors than any prompt.

Email threads

Email threads have a specific problem: the newest message is on top, quoted history repeats in every reply, and the actual question is buried in message six. Paste the whole thread and ask for the timeline first:

"Reconstruct this thread as a dated timeline: who said what, in order. Then tell me what is currently being asked, of whom, and what has already been answered. Finally, list anything promised by anyone that has not been delivered."

The timeline forces the model to untangle the quoting, and it lets you see whether it understood who is who before you trust the rest.

Checking for omissions

Every summary, structured or not, gets this second prompt:

"Go back to the source. List every number, date, proper name, and commitment that appears in the document and is not in your summary. Then tell me which of those, if any, would change my reading of the document."

You will be surprised how often the list is not empty. Then do two manual checks that take a minute each: read the source's last two pages (models drop the end more than the beginning), and search the source for one term you care about (a deadline, a dollar amount, a client name) and confirm it appears in the summary in the same form.

For high-stakes documents, add a third: ask a second model, or the same model in a fresh chat, to summarize independently, and compare. Disagreement between two summaries is your reading list.

The fact-checking guide has the fuller routine, and the hallucinations guide explains why fluent summaries can be confidently wrong.

Grounded Q&A: tools that cite your files

For documents you will come back to, a one-off summary is the wrong shape. You want to ask questions and get answers that point to the passage they came from.

NotebookLM

NotebookLM from Google is built for this. You add sources (PDFs, Google Docs, pasted text, web pages, and more), and every answer comes with numbered citations that open the exact passage. It answers from what you uploaded, which makes it far less likely to introduce outside "facts," and it will say when the sources do not cover a question. It can also generate briefing documents, FAQs, and audio overviews from the same sources; treat those as drafts and check them like any other summary.

A good pattern: create one notebook per project or client, add every relevant document as it arrives, and ask it the three questions from earlier in this guide before every meeting.

File upload in the main assistants

ChatGPT, Claude, and Gemini all accept uploaded PDFs, Word files, and spreadsheets, and their project or folder features let you keep a set of documents attached across many chats. Microsoft Copilot can summarize files in OneDrive and SharePoint and threads in Outlook without leaving those apps, which is often the right choice inside a company that has licensed it.

Whatever tool you use, ask for citations and click through. Grounded tools reduce invention; they do not eliminate misreading.

Worked example: a 90-page vendor proposal

The situation: an operations manager receives a 90-page proposal from a warehouse-management software vendor, needs a recommendation by Friday, and has a specific question: total three-year cost, what is excluded, and what happens at renewal.

She uploads the PDF to her company's approved assistant and runs the structured prompt with that question at the top. Section 4 of the output (numbers and dates) lists a license fee, an implementation fee, and a "professional services estimate," each with a page reference. Section 6 (conditions) flags a sentence on page 61: the implementation fee assumes "clean data provided by the customer."

She runs the omission check. It returns one item: an annual uplift clause on page 74 that the first pass missed. That clause changes the three-year number materially.

She reads pages 61 and 74 herself, builds the three-year total in a spreadsheet from the cited figures, and writes a one-page recommendation. Total time: about an hour, against a full day of reading. The proposal was not read cover to cover; it was read where it mattered, and the model told her where that was.

Confidentiality

Documents are where confidential material lives: contracts, board decks, patient records, personnel files, client matters. The rule: never paste confidential data into a consumer AI tool unless your organization has approved it. Use the enterprise version your IT team has vetted, or the AI built into your document system, and check whether the tool trains on your inputs (most business tiers say no; consumer tiers vary).

For regulated roles the stakes are higher. Health information falls under HIPAA. Legal documents may carry attorney-client privilege that an outside tool could compromise. HR files contain personal data (PII). When in doubt, strip identifying details before uploading, or do not upload.

Next steps

• Copy ready-made prompts from meetings and summaries.
• Read the tool page for NotebookLM and the glossary entries on context window, summarization, and grounding.
• Learn the full verification routine in how to fact-check AI output.
• See how project managers and paralegals apply this to their own document loads.
• Improve the prompts themselves with the prompt engineering guide and the Chain of Density technique in prompt frameworks that work.
• The course has a research and knowledge-work module built around these methods.