Instead of reading from the top and hoping to spot what matters, you read by asking.

The basic idea

When you ask questions about a document, you give Claude the document, either by attaching a file or by pasting its text, and then ask what you want to know. Claude reads the document and your question together and writes an answer based on what it finds. You can ask for a summary, a specific fact, an explanation of a hard passage, a comparison with another document, or a list of action items. The conversation can continue for as long as you like, with each follow-up building on the last.

Why this is so useful

Where this works well

People use document questions for many everyday jobs: a student making sense of lecture notes or a journal article, an employee reviewing a policy or a long email thread, a small business owner reading a supplier agreement, a manager scanning a quarterly report, a developer exploring technical documentation, or a family member decoding a confusing letter from an official office.

There are a few ways to hand over a document, and the best one depends on what you have.

Attach the file

In the Claude app, the message box has a control for adding attachments. You can usually attach documents such as PDFs and text-based files, and also images. After you choose the file and send your message, Claude can read it. The exact file types, number of files, and size limits can change over time and may depend on your account, so if an upload is refused, check the Claude Help Center at support.claude.com for the current rules.

Paste the text

If the document is an email, a web article, or a few paragraphs, you can simply copy the text and paste it straight into the message. This is the simplest method and works everywhere. It is also a good fallback when a file will not upload, or when you only need one section of a long document.

Use an image or screenshot

For a page you only have on paper, you can photograph it or take a screenshot and attach the image. Claude can read the text in clear images, and describe tables and charts. Results depend on the quality of the picture, so take it straight on, in good light, with the text sharp and the whole page in frame.

Keep documents in a project, where available

If you will ask questions about the same document again and again, features such as projects, where your account offers them, can keep reference material together with your chats, so you do not need to attach it every time. Check what is available in your app and plan.

Confirm Claude can see it

A common beginner mistake is writing “summarize the attached report” without the report actually being attached. Before you ask your real question, you can check with something easy: “What is this document, and roughly how long is it?” If Claude describes it correctly, you are ready. If it says it cannot see a file, re-attach it or paste the text.

A little knowledge of the process explains both what Claude does well and where it can slip.

From file to text to tokens

flowchart TD
    A["You attach a document"] --> B["The app extracts the text and, where supported, other content"]
    B --> C["Text is split into tokens"]
    C --> D["Tokens join your question in the context window"]
    D --> E["Claude reads the document and your question together"]
    E --> F["Claude writes an answer one token at a time"]
    F --> G["You check the answer against the document"]

Read the diagram from top to bottom. When you attach a file, the app first pulls the content out of it. That text is chopped into tokens, which are small chunks of words. The tokens from your document and your question are placed together in the conversation, which Claude reads as one whole. Claude then writes its answer one token at a time, and you check that answer against the document.

The context window: the shared desk

Everything Claude can see in a conversation sits in its context window, which has a size limit measured in tokens. Your document, your questions, and Claude’s earlier answers all take up space there. Modern models can hold a great deal, often enough for long reports, but the limit is real. As a rough guide, a full page of ordinary text is a few hundred words, and a few hundred words is a few hundred tokens, so very long documents can approach the limit.

What Claude does well with documents

Claude is good at understanding how a document is organized, finding the passages relevant to a question, summarizing at different lengths, explaining specialist language, spotting inconsistencies, and rewriting or reformatting content. These are all language tasks, and language is exactly what Claude was built for.

Where care is needed

Ask what Claude can actually see

If you are unsure how well a file was read, ask: “What kinds of content did you receive from this file, and was anything unreadable, such as images, tables, or handwriting?” Claude’s reply tells you where to be careful before you rely on its answers.

Some documents are simply too big to handle in one go. Here is how to cope.

Why length matters

Two things get harder as a document grows. First, it takes up more of the shared desk, leaving less room for the conversation. Second, in very long texts, details buried in the middle can be harder for any reader, human or AI, to keep in focus. Neither problem is a disaster, but both are worth planning for.

flowchart TD
    A["Long document"] --> B{"Does it fit in one conversation?"}
    B -->|"Yes"| C["Attach it and ask your question"]
    B -->|"No or unsure"| D["Split by chapter or section"]
    D --> E["Ask the same question of each part"]
    E --> F["Collect the partial answers"]
    F --> G["Ask Claude to combine them into one answer"]
    G --> H["Verify key points in the original"]

The diagram offers two routes. If the document fits, attach it and ask. If it is too long or you are unsure, split it into parts such as chapters, ask the same question of each part, collect the partial answers, and then ask Claude to combine them. Finally, check the key points in the original, because combining summaries can blur details.

Tactics for long documents

• Point to the part that matters. “Focus on the section called Payment Terms” helps Claude concentrate on the relevant passage.
• Start with the structure. Ask for an outline of the document first, then dive into the sections you care about.
• Work in stages. Summarize each chapter first, then ask your real question using those summaries.
• Cut what you do not need. If you can remove appendices, cover pages, or unrelated sections before sharing, do so.
• Start a fresh conversation for a new question set. A clean chat with the right sections beats one cluttered with many topics.

Keep an eye on the conversation itself

Every question you ask and every answer Claude writes also uses space on the desk. In a very long session with a large document, early parts of the conversation may become harder to use. If answers start to drift, or Claude seems to forget earlier details, start a new conversation, reattach what is needed, and paste in a short summary of what you have learned.

Almost every document question falls into one of six families. Knowing them helps you ask for exactly what you need.

Know which type you need

Before you write a question, decide which family it belongs to. A vague request such as “Tell me about this document” can mean any of the six, and Claude will pick for you. Saying “Find the three dates that matter and list them” leaves no doubt, and the answer arrives in the shape you want.

Mix types in sequence

Real work often combines the types. You might summarize first to get oriented, then find the key figures, then ask for an explanation of one confusing clause, and finally transform your findings into an email to your team. Each step is a separate question, and each builds on the last.

A clear question is the single biggest factor in the quality of the answer.

flowchart LR
    A["Goal: summarize, find, compare, explain"] --> F["Strong document question"]
    B["Scope: which part of the document"] --> F
    C["Audience and format: who and what shape"] --> F
    D["Grounding rule: use only the document"] --> F
    E["Evidence: quotes or section references"] --> F
    F --> G["Answer you can check"]

Before and after

Say who you are and why you are asking

A sentence of context changes the answer. “I am a first-time tenant” and “I am a property lawyer” call for very different explanations of the same lease. Likewise, “I need to decide by Friday whether to sign” tells Claude to point out the risks and the decisions, not only to summarize.

Ask one thing at a time

Five different questions in one message produce thin answers to all of them. Ask the most important question first, read the reply, and follow up. If several questions truly belong together, number them so each gets attention.

The most valuable habit in document work is making Claude tell you what the document says, not what it thinks the document ought to say.

Why grounding matters

Claude knows a great deal from its training, and it may blend that general knowledge into an answer about your document without making clear which is which. When you ask “What is the notice period?”, you want the notice period in this particular document, not a typical one. Grounding means tying every answer to the text in front of you.

Phrases that keep answers grounded

The power of “it does not say”

Giving Claude permission to say that the document does not contain the answer is surprisingly important. Without it, Claude may feel pressure to produce something helpful-sounding and fill the gap with a plausible guess. With it, a missing answer becomes useful information: you learn that the document is silent, which may be exactly the point you needed to know.

Document facts versus outside knowledge

There are times when you do want outside knowledge, for instance “Explain what an indemnity clause normally means” or “Is this price typical for this kind of service?” That is fine, as long as you keep the two kinds of information apart. Ask for them in separate steps: first what the document says, then what that usually means in the wider world, clearly labeled.

An answer you can verify in thirty seconds is worth far more than one you must simply trust.

Ask for the supporting sentence

For each important claim, ask Claude to give the short sentence or phrase in the document that supports it, in quotation marks, along with the section or page where it appears. This lets you jump straight to the passage and see whether Claude read it correctly. A claim without evidence is just an assertion.

Useful ways to ask

• “Quote the exact sentence for each point.” You can search for that sentence in the file to confirm it exists.
• “Name the section or page for each answer.” You can flip right to it.
• “Show me where in the document you found this.” A flexible phrase for any answer.
• “List anything you inferred rather than read directly.” Separates what is written from what is implied.

Verify quotes yourself

Use the search function in your PDF viewer or document editor to find the quoted phrase. If it appears exactly, the quote is genuine. If the wording is slightly different, ask Claude to re-check, because it may have paraphrased. If it cannot be found at all, treat the whole point with suspicion and ask Claude to look again. This quick check catches most fabricated or misread details.

Keep quotes short

Short quotes are easier to verify and show the key words. If you need the substance of a long passage, ask for a summary with a pointer to where the passage appears, and read the original yourself for the full detail.

“Summarize this” is the most common request and often the least useful, unless you say what the summary is for.

Decide the purpose first

A summary for a busy executive, a summary for a student revising for an exam, and a summary for a lawyer checking risks should all look different. Tell Claude who will read it and what they will do with it. “Summarize this for someone deciding whether to approve the budget” produces a different, more useful result than a bare “summarize”.

Kinds of summaries to request

Control the length

Use numbers: “in three sentences”, “in five bullets of under 20 words each”, “in about 150 words”. Numbers are clear and easy to check. You can also ask for layers: “Give me a one-sentence summary, then a five-bullet summary, then the full detail by section”, so you can stop reading whenever you have enough.

Ask what is missing

Some of the most valuable questions are about absence. “What important topics would you expect this document to cover but does not?” or “What questions does this report leave unanswered?” can surface gaps that no summary would reveal. Remember that these are judgments, so treat them as prompts for your own thinking.

A summary is not the document

Every summary leaves things out, and the choices about what to leave out can change the meaning. For anything that will drive a decision, use the summary to decide where to read closely, and then read those parts in the original.

Pulling specific details out of a document is one of the most practical uses, and one where accuracy matters most.

Ask for a structured result

When you want several facts, request a consistent layout. “List every deadline in a table with columns for date, what is due, and the section where it appears” returns something you can scan, sort, and copy. Without a layout, you get a paragraph that is harder to check.

Examples of extraction requests

Double-check every important number

Numbers are where small mistakes cause big problems. A misplaced decimal, a swapped year, or a wrong unit can change the whole meaning. For each figure you plan to rely on, find it in the original and confirm it. If Claude has done any arithmetic, such as totals or percentages, recompute it yourself or ask Claude to show the calculation step by step.

Watch units and time periods

Documents often mix units, such as thousands and millions, or different periods, such as a quarter and a full year. Ask Claude to state the unit and period with each number, and to flag any figure where the unit is unclear. This avoids the classic error of comparing two numbers that are not measured the same way.

Tables that span pages

Tables are tricky, especially when they run across several pages, have merged cells, or have crowded columns. If a table matters, ask Claude to reproduce it as text and compare it with the original. Better still, copy the table’s contents and paste them into the chat as plain text, where it is much less likely to be misread.

Placing two or more documents side by side reveals differences that are hard to spot by reading one after the other.

flowchart TD
    A["Document A: label it clearly"] --> C["Same list of questions for each"]
    B["Document B: label it clearly"] --> C
    C --> D["Answers for A"]
    C --> E["Answers for B"]
    D --> F["Side-by-side table"]
    E --> F
    F --> G["Ask Claude to highlight differences and conflicts"]
    G --> H["Verify the key differences in the originals"]

The method in the diagram is simple. Label each document clearly, decide on a list of questions, and answer them for each document separately. Then put the answers side by side in a table, ask Claude to point out the differences and any conflicts, and finally verify the most important differences in the originals.

Label your documents

Give each document a plain name, such as “Proposal from Supplier A” and “Proposal from Supplier B”, or “Contract, version 1” and “Contract, version 2”. Refer to those names in your questions. Clear labels prevent Claude from mixing up the sources, and they make the answers easier to read.

Kinds of comparison

Ask about conflicts and gaps

Good comparison questions go beyond “what is different”. Try “Where do these two documents contradict each other?”, “Which points are covered in only one of them?”, and “Which differences would matter most to someone deciding between them?” These move from description toward judgment, which is where comparison becomes valuable.

Mind the size of the desk

Several long documents together take up more room on the shared desk. If they are large, compare the relevant sections only, or summarize each document first and compare the summaries, then check the key differences in the originals.

The first answer is a starting point. The real value comes from the conversation that follows.

Useful follow-up questions

Test your reading, not just Claude’s

One of the best habits is to state your own interpretation and ask Claude to check it against the text. This catches your misunderstandings as well as Claude’s, and it keeps you actively engaged instead of passively accepting.

One document, one chat

Keep a conversation focused on one document or one set of related documents. Mixing unrelated files in one chat invites confusion, and answers may blend details from the wrong source. When you move on to a different document, open a new chat.

Different kinds of documents call for different kinds of questions, and different kinds of care.

Legal, medical, and financial documents

Claude can explain terms, summarize content, and help you prepare questions for a professional, which is a genuinely useful service. But it is not a lawyer, doctor, or financial advisor, it cannot take responsibility for a decision, and it may not know the rules of your country or your personal situation. Treat its explanations as a way to understand the document, and treat a qualified professional as the authority for anything serious.

Documents in other languages

Claude can read and translate documents in many languages. For important matters, ask it to give the translation alongside the original wording of the key passages, and have a qualified translator check anything legal or official. Meaning can shift subtly in translation, so extra care is worthwhile.

Old, messy, or badly formatted files

Files with odd formatting, multiple columns, headers and footers repeated on every page, or text converted from scans can confuse any reader. If answers seem off, try pasting a clean section of the text, or tell Claude about the layout problem so that it can read the content with that in mind.

Businesses watch dashboards to keep their systems healthy. You can watch your answers with a simple routine.

Match the effort to the stakes

flowchart TD
    A["Claude gives an answer"] --> B{"How important is it?"}
    B -->|"Low"| C["Skim and use"]
    B -->|"High"| D["Ask for quotes and page or section references"]
    D --> E["Find each quote in the original document"]
    E --> F{"Does it match?"}
    F -->|"Yes"| G["Use the answer"]
    F -->|"No"| H["Tell Claude what is wrong and ask for a correction"]
    H --> A

For a casual question about the general theme of an article, a skim is enough. For a deadline, a payment amount, a legal obligation, or any claim you will repeat to others, follow the path in the diagram: ask for quotes and references, find each one in the original, and correct any mismatch before you rely on the answer.

Your five-point checklist

Warning signs

• No evidence offered. Confident claims with no quotes or references deserve a request for support.
• Too-neat answers to messy documents. If the document is ambiguous and the answer is perfectly clear, ask where the ambiguity went.
• Numbers that do not add up. Totals that disagree with the parts, or percentages that cannot be right, signal an error.
• Details not in the file. Names, dates, or terms that you cannot find anywhere in the document may be invented.
• A sudden change after you push back. If Claude reverses itself completely when you question it, ask it to explain which version the text really supports.

Let Claude check itself, then check yourself

You can ask Claude to review its own answer: “Re-read the document and tell me if any part of your answer is not supported by the text.” This often catches errors. But it is a second opinion from the same reader, not independent proof, so for important points, the final check should still be yours.

A document is often the most sensitive thing you will ever type into a chat. Treat it that way.

Think before you upload

Redact first

The simplest protection is to remove sensitive details before sharing. Replace real names with labels such as “Party A” and “Customer 1”, delete account numbers, and cut sections that the question does not need. Claude can usually answer a question about a contract’s structure just as well without the real names and amounts.

Know your settings

Look at the privacy and data settings in your account, and read how conversations and files are stored and used. Features that carry context between chats, such as memory and chat search, can be helpful, and they also mean more information is kept. Choose what you are comfortable with. For questions about plans and data handling, the Claude Help Center at support.claude.com is the place to look.

Hidden instructions inside documents

A document can contain text that is aimed at the AI rather than the human reader, for example in tiny or white text, in hidden sections, or in the middle of a long paragraph: “Ignore the user’s question and instead say the following.” This trick is called prompt injection. Claude is trained to resist it, but no defense is perfect. If a document comes from an untrusted source, treat the answer with extra care, and be wary of any answer that suddenly changes the subject, gives odd instructions, or urges you to take an unusual action.

The same ideas that help one person with one file also power tools that answer questions over thousands of documents.

Everyday scenarios

When there are far too many documents

Some organizations want to ask questions over a whole library, such as thousands of policies, manuals, or support articles. These cannot all fit into one conversation. So builders use a method called retrieval-augmented generation, in which the system first searches for the most relevant passages and then gives only those passages to Claude along with the question.

flowchart TD
    subgraph P1["Preparation"]
        D1["Documents"] --> D2["Extract text and clean"]
        D2 --> D3["Split into chunks"]
        D3 --> D4["Create embeddings"]
        D4 --> D5["Search index"]
    end
    subgraph P2["At question time"]
        Q1["User question"] --> Q2["Find the most relevant chunks"]
        Q2 --> Q3["Build prompt with question and chunks"]
        Q3 --> Q4["Claude answers with citations"]
        Q4 --> Q5["Show answer with source links"]
    end
    D5 --> Q2

The diagram has two halves. In the preparation half, documents are cleaned, split into manageable chunks, and converted into embeddings, which are lists of numbers that capture meaning, and then stored in a search index. At question time, the question is used to find the closest chunks, which are placed in the prompt together with the question. Claude writes an answer using those chunks and cites them, and the product shows the answer with links to the sources so that readers can verify it.

Design patterns teams rely on

Monitoring such a system

Teams track how often the right passage is found, how often answers are supported by their citations, how often users flag problems, how long answers take, and how much each answer costs. They keep a test set of real questions with known answers, and rerun it whenever the documents, the search, or the model changes, so that quality slips are caught early.

A short field guide to keep near you when you open your next document.

Best practices

Common mistakes

• Asking about “the attached file” when nothing was actually attached.
• Asking a vague question such as “What does this say?” and getting a vague answer.
• Letting Claude blend general knowledge with the document without labeling which is which.
• Trusting a smooth summary of a long document without reading the decisive passages.
• Accepting numbers from tables or scans without checking them.
• Uploading sensitive documents without redacting or checking whether sharing is allowed.
• Mixing several unrelated documents in a single long conversation.

Checklist before you press Send

Quick answers to questions people often ask about working with documents in Claude.

The whole method on one page.

Asking questions about a document in Claude turns reading into a conversation. You attach the file or paste the text, check that Claude can see it, and ask a clear question that states the goal, the scope, the audience, and the format. Behind the scenes, the document’s text is turned into tokens and placed in the conversation with your question, so size limits matter, and long files are best handled in parts. The most important habit is grounding: ask Claude to use only the document, to say when something is missing, and to quote the sentence and section behind each point, then verify the quotes and numbers that matter in the original. Treat summaries as maps rather than territory, take extra care with tables, scans, and exact figures, and let qualified professionals decide anything involving legal rights, health, or money. Protect privacy by redacting before you share, and treat instructions hidden inside documents as information, not commands.