Mindket

How to summarize a PDF with AI and check the summary

A black banner titled "AI at work" over the line "Emails, research, documents and spreadsheets."

To summarize a PDF with AI, upload it to a chat app, name the reader and the shape you want, then ask for the quotes and page numbers it used. Check a few claims against those pages. Each app has file limits, so split long documents.

In this lesson

  • Attach the file and name the reader of the summary.
  • Ask the app to quote its sources with page numbers, then test three claims.
  • Know your app's file limits, and split a long document.
  • Keep private material out of an upload unless your company allows it.
All courses
On this page
  1. Before you start
  2. Summarize a document in five steps
  3. File limits
  4. Meeting notes work the same way
  5. Before you upload something private
  6. How Andrej Karpathy reads with an LLM in three passes
  7. Common mistakes
  8. Try it yourself
  9. Next: spreadsheets
  10. Check yourself
  11. Frequently asked questions
  12. Sources

Before you start

You need a document you may share with an AI app, and a chat app that takes files. This lesson follows Claude, Gemini and Copilot in Word, because their help pages state their limits. The steps are alike in other apps.

A PDF made from a word processor holds real text. A scan is a picture of pages, so a summary of one is the case to check hardest. If you can, start from the original file.

Summarize a document in five steps

The prompt below carries steps two to four. It is our own sample, so change the reader and the shape to fit your document.

Anthropic's advice for long documents

For long document tasks, ask Claude to quote relevant parts of the documents first before carrying out its task. This helps Claude focus on the relevant content and ignore the rest of the document.

Anthropic, prompting guide for Claude · Anthropic: Prompting best practices (read November 4, 2026)
Summarize the attached report for my manager, who has not read it and will skim it on her phone.

Give me:
1. Three bullet points, each under 25 words.
2. The one decision the report asks for.
3. Any numbers a reader could challenge, each with its page number.

Before you summarize, quote the sentences you rely on.
If something is not in the document, say "not in the document".
  1. Attach the file. In Claude, click the + button at the lower left of the chat box, choose the option to add files or photos, and pick the file. You can also drag it into the window. In Gemini, enter your prompt at gemini.google.com, click Add files, choose Upload files, then select Submit. In Word, ask Copilot Chat to summarize the document.
  2. Say who the summary is for. "For my manager, who will skim it on her phone" tells the app how short and how plain to write.
  3. Describe the shape. Give the count and the length, such as three bullets under 25 words, and name the one thing you most need, such as the decision the report asks for.
  4. Ask for quotes and page numbers. Quotes let you find the source sentence in seconds. When you point to a page in a prompt, use the number your PDF viewer shows, not the number printed on the page. That is Claude's tip.
  5. Check three claims against the pages. Pick the numbers and the names a reader could challenge. If a quote is not on the page, ask the app to look again, and trust nothing else from that answer until it does.

File limits

Limits are why an upload fails or a summary comes back thin. The table gives the figures the three help pages state on 5 November 2026, and plans can change them.

A long document is where summaries go wrong. Google's page says a file that is too large can make Gemini miss details, and Claude's page tips you to split large documents into smaller sections. Summarize each part, then ask for a summary of the summaries.

Claude's upload page, Google's upload page and Microsoft's page on Word summaries, read on 5 November 2026.
ClaudeGeminiCopilot in Word
Files at onceUp to 20 per chatUp to 10 per promptThe document you have open
File size500 MB per file in a chat, 30 MB per file in a project100 MB per file, and 2 GB for a videoNot stated on the page
LengthPDFs up to 1,000 pages. Up to 100 pages it reads text and visuals such as charts, and beyond that text onlyA 1 million token window covers about 1,500 pages of text. Past the window, answers can miss detailsAn automatic summary needs at least 200 words
File typesPDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB, JSON and XLSXMost file typesA Word document
Watch forA document that is not a PDF is read as text only, so embedded images are skippedIf it fails to read a file, upload it againIt reads the whole document but does not always cite later parts
Pages of text one file can bring inClaude's two bars are its page limits for one PDF. Gemini's bar is Google's estimate of what its 1 million token window holds, not a page limit.
  • Claude, with charts and images read100 pages
  • Claude, text only1,000 pages
  • Gemini, text in its context window1,500 pages

Sources: Anthropic: Uploading files to Claude (read November 5, 2026), Google: Upload and analyze files in Gemini Apps (read November 5, 2026)

Word's automatic summary also needs a Microsoft Copilot work license and a document saved in OneDrive or SharePoint. Without them you can still ask Copilot for a summary.

Meeting notes work the same way

Paste the notes into the chat, or save them as a text or Word file and attach it. The ask is different: you want decisions, owners and dates, and you do not want the app to guess any of them. Tell it what to write when something is missing.

Below are my notes from today's team meeting.

List three things:
1. Decisions we made.
2. Tasks, each with an owner and a due date.
3. Open questions.

If an owner or a date is missing from my notes, write "missing" instead of guessing.

Notes:
[paste your notes here]

Before you upload something private

A contract or a customer list is not a casual upload. The rows below are what Google and Anthropic say about their consumer apps on 4 November 2026. They describe consumer accounts. If your employer gives you a work account, ask your IT team what applies before you upload company files.

Google's warning, word for word

Please don’t enter confidential information that you wouldn’t want a reviewer to see or Google to use to improve our services, including machine-learning technologies.

Google, Gemini Apps help · Google: Gemini Apps Privacy Hub (read November 4, 2026)
App and settingWhat the help page says
Gemini, Keep Activity onChats and uploads are saved in your Activity and used to improve services, including training generative AI models
Gemini, Keep Activity offFuture chats are not used to train the models unless you send feedback, and they are kept with your account for 72 hours
Gemini, human reviewA subset of chats is reviewed by people, and reviewed chats are kept for up to three years even if you delete your activity
Claude Free, Pro and MaxChats are used for training only if you opt in to Model Improvement, a chat is flagged for safety review, or you opt in another way. Incognito chats are not used for training

How Andrej Karpathy reads with an LLM in three passes

Andrej Karpathy was a founding member of OpenAI and Tesla's director of AI from 2017 to 2022, as his own site lists his roles. On 18 November 2025 he posted on X that he had started reading everything with LLMs: blogs, articles, book chapters. He reads the piece himself first. The second pass asks the model to explain or summarize it, and the third is questions and answers.

He wrote that he usually comes away with a deeper understanding than if he had moved on, and that it was becoming one of his top uses. This lesson's five steps start at his second pass. For a report you must act on, put his first pass back. Skim it yourself before you ask, and you will notice when a summary drops the part that caught your eye.

His third pass is where quotes and page numbers pay off. Hold every follow-up question to the rule the summary followed: an answer, and the sentence it came from.

Usually pass 1 is manual, then pass 2 “explain/summarize”, pass 3 Q&A.

Andrej Karpathy, founding member of OpenAI and former director of AI at Tesla · X (read November 5, 2026)
Andrej Karpathy (@karpathy) on X, 18 November 2025: I am starting to get into a habit of reading everything (blogs, articles, book chapters) with LLMs. Usually pass 1 is manual, then pass 2 explain or summarize, pass 3 Q&A. I usually end up with a better, deeper understanding than if I moved on. Growing to among top use cases.
Andrej Karpathy on X, 18 November 2025

A screenshot of X's own embed of the post, taken on 5 November 2026, with the time in UTC. X cuts the post short with Show more. Select it to open the post on X.

Common mistakes

  • Trusting a summary you have not checked. Open three of the pages it cites and find the sentences.
  • Uploading a file far past the limits and wondering why the answer is thin. Split it first.
  • Asking for "a summary" with no reader and no shape. You get a generic one.
  • Trusting the summary of a scan. A scan is the case where checking matters most.
  • Uploading a private file to a consumer app because it was quicker. The rules above come first.

Try it yourself

  • Take a report you have already read and ask for the summary with quotes and page numbers.
  • Compare it with what you remember. Mark anything the summary left out or got wrong.
  • Run your last meeting notes through the notes prompt and see where it writes "missing".
  • Find your app's file limits in the table and check whether your longest document fits.

Next: spreadsheets

Documents hold words. The next lesson, how to analyze Excel data with AI, handles numbers, where a wrong answer is easier to miss and the check is different.

The habit of asking for quotes comes from the same idea as giving a model context: what you put in front of it is what it works from. What are tokens in AI explains the units behind a window's size.

Check yourself

  1. What should you ask for so you can check a summary?

    Show answer

    C. The quotes and page numbers it relied on Anthropic says to have the model quote the relevant parts first, and page numbers let you find each quote.

  2. Claude is given a PDF of 150 pages. What does it read?

    Show answer

    A. Text only Claude reads text and visuals up to 100 pages and text only from 101 to 1,000 pages. Over 1,000 pages the upload fails.

  3. Google's page says a file that is too large can cause what in Gemini?

    Show answer

    D. Gemini misses details Google says large files can make Gemini miss details, so use smaller files.

  4. Which line in a meeting-notes prompt keeps the app from guessing a task owner?

    Show answer

    B. If an owner or a date is missing, write "missing" Telling the app what to write when information is absent keeps it from inventing an owner or a date.

Frequently asked questions

Can AI summarize a scanned PDF?

Sometimes. The help pages do not promise that every scan can be read, and Claude's page says it reads the text and visuals of PDFs up to 100 pages. Ask for a quote from the middle of the file and compare it with the page.

How long a PDF can AI summarize?

Claude counts pages and Gemini counts file size, so the two answers differ. The table above has both. Anything near either limit is better split into parts.

What do I do when the document is longer than the limit?

Split it by chapter or section, summarize each part with the same prompt, then paste the part summaries into a new chat and ask for one summary of them. Claude's page also suggests splitting large documents.

Is it safe to upload a confidential contract to an AI app?

Not by default. On the consumer apps, Google says chats can be reviewed by people and used to improve services, and Anthropic says chats are used for training if you opt in. A work account may follow other terms, so ask before you upload.

Can Word summarize a short document?

Yes. Microsoft says automatic summaries need at least 200 words of content, but for a shorter document you can still select Summary at the top of the page to generate one.

Can I summarize several documents at once?

Yes. Attach them together and ask for one summary per file first, then one that combines them. Both apps cap the number of files, as the table shows.

Sources

  1. Anthropic: Uploading files to Claude (read November 5, 2026)https://support.claude.com/en/articles/8241126-uploading-files-to-claude
  2. Google: Upload and analyze files in Gemini Apps (read November 5, 2026)https://support.google.com/gemini/answer/14903178?hl=en
  3. Microsoft: Create a summary of your document with Copilot in Word (read November 5, 2026)https://support.microsoft.com/en-us/office/create-a-summary-of-your-document-with-copilot-79bb7a0a-3bf7-41fe-8c09-56f855b669bf
  4. Google: Gemini Apps Privacy Hub (read November 5, 2026)https://support.google.com/gemini/answer/13594961?hl=en
  5. Anthropic: Consumer model training (read November 5, 2026)https://privacy.claude.com/en/articles/10023580-consumer-model-training
  6. Anthropic: Prompting best practices (read November 4, 2026)https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices
  7. Andrej Karpathy (@karpathy) on X: reading with LLMs (read November 5, 2026)https://x.com/karpathy/status/1990577951671509438
  8. Andrej Karpathy: personal site (read November 5, 2026)https://karpathy.ai/

Share this lesson

See how a workflow reads files for you

Mindket lists automation workflows you can open and read before you use them: their nodes, the apps they connect and their setup steps.

Browse automations

This lesson as Markdown · AI at work: a free course for beginners · All courses