Analyze a long PDF with Claude: a page-cited workflow

Long-document analysis works best when you ask Claude to build a map of the document before asking for conclusions.

1. Check the source file

Confirm that the PDF contains readable text, page numbers, and complete appendices. OCR scanned pages before analysis.

2. Request a document map

Ask for the purpose, section structure, key entities, dates, definitions, and a list of tables or exhibits.

3. Ask narrow questions first

Work section by section. Require page references and ask Claude to separate direct evidence from interpretation.

4. Build a contradiction table

List claims that conflict across sections, changes in definitions, unsupported assumptions, and missing data.

5. Create the final synthesis

Only after verification, request a concise summary organized around the decision you need to make.

Quality check

Open every cited page for the most important claims. If a citation is wrong, re-check the surrounding section rather than accepting a corrected sentence automatically.

Practical workflow

Use a long-document workflow that preserves page references and separates extraction from interpretation. The steps below are designed to leave a human-readable record of what was tried, what was approved and what remains uncertain.

Confirm the file is permitted to upload and remove unnecessary sensitive pages.
Ask for a page-referenced outline before requesting conclusions.
Extract claims, numbers and definitions into a review table.
Check important passages in the original PDF, including footnotes and appendices.
Keep the final summary beside the source version and date it.

Worked example

For a policy PDF, ask for a table of requirement, page, exception and implementation owner. If the assistant cannot point to the page, treat the statement as a lead to verify, not a final answer.

Decision table

Checkpoint	Pass condition	Stop condition
Scope	The task, audience and allowed data are explicit.	Inputs or approval owner are unclear.
Evidence	The important claim or change points to a source, test or artifact.	The result depends on an unsupported assertion.
Review	A named person can reproduce the check and explain the decision.	The output is accepted only because it sounds fluent or looks polished.
Handoff	Files, settings, rights and rollback steps are recorded.	The next operator cannot tell what changed.

Limits and sources

Adapt this method to the tool, data sensitivity and risk of the task. It is a practical editorial workflow, not a claim of independent product performance.

Official sources to verify: support.anthropic.com, help.openai.com.

Continue with: Claude, Chatgpt, How To Review Ai Tool Privacy.

Best fit and limitations

Best fit: analysts, students, and teams who regularly work with long documents and can define the questions and source boundaries in advance. Not fit: legal, medical, or financial decisions that depend on one unverified extraction. Our editorial assessment is that long-context convenience matters only when the reviewer checks citations, page references, and missing context.

Limitations: PDFs with scans, tables, footnotes, or mixed languages may extract unevenly; plan limits and file handling can change. Use the Claude review with a small representative sample first, then follow the privacy review guide before uploading sensitive files.

November 2026: Verify PDF page limits and visual analysis mode

Anthropic’s current Claude.ai help page distinguishes visual PDF analysis from text extraction. It says Claude 4 models, Claude 3.7 Sonnet, and Claude 3.5 Sonnet can analyze both text and visual elements in PDFs that are under 100 pages when uploaded to chats or projects. For PDFs over 100 pages, Claude processes text only; text-only handling also applies when using other models. The guide title should therefore not be read as a guarantee that a 100-page file receives visual analysis or that every page is understood without verification.

The same help page lists a 30 MB per-file upload limit and up to 20 files per chat. Additional token limits also depend on extracted content and context. These are current product limits stated by Anthropic and may change. If a scanned or image-heavy PDF matters, confirm that the account and model support the needed visual mode, and test the file before relying on charts, diagrams, or page layout.

Safer workflow:

Check whether the file is text-based or scanned, and confirm the PDF page count in a local viewer.
If the document is 100 pages or longer, divide it into clearly named sections if visual details matter; do not assume a text-only extraction preserves chart meaning.
Ask for a document map first, then request a table of claim, PDF-viewer page number, supporting passage, and uncertainty.
Open the cited page in the original file. Check footnotes, appendices, charts, table headers, and any conflicting definition before using a conclusion.
Keep the final answer next to the source file and date it. If a page reference cannot be found, treat the statement as a research lead, not a verified result.

Decision impact: Claude can help organize a long document, but the right workflow depends on whether the task needs text retrieval or visual interpretation. For a contract, policy, report, or research paper, page-cited extraction is useful; it does not replace a qualified legal, financial, medical, or domain review. Do not upload confidential material unless the account’s data handling and organizational approval have been checked. See our source-verified AI research workflow and AI privacy review guide.

Our assessment: the original “100-page” framing was too broad because Anthropic draws a specific boundary for visual processing at fewer than 100 pages and distinguishes text-only processing. The revised title and instructions should make that limitation visible before a reader chooses a model or uploads a file. AI Tools Manual did not benchmark PDF extraction accuracy.

Primary source: Anthropic’s official document-upload and PDF-processing guidance. Next review: November 30, 2026; verify page-count, file-size, model, and account limits before reuse.