Subscribe
Accounting & FinanceTested · Oct 9, 2026Claude Pro (Claude desktop app

How to Analyze Long Documents with Claude: Annual Reports, Contracts and Filings

by Waheed Burna | Oct 13, 2026 | Accounting & Finance, Data & Research

A five-step method for getting reliable answers out of 200-page reports, leases and agreements with Claude: map the structure, extract facts with page references, ask cited questions, summarize in two stages and check the result.

i
TL;DR Summary

Claude can read a typical annual report in one go, but a one-pass summary of a long file is unevenly wrong in ways you can't see. Map the structure first, pull figures and obligations out with page numbers, ask questions that come back with a page reference after every sentence, and summarize section by section before writing a briefing. In PDFs over 100 pages Claude reads text only, so check charts and scanned tables yourself.

!
The Bottom Line

Treat Claude as a fast, thorough reader that must show its sources. Make it cite pages, summarize in two stages, start a new conversation for each new question, and check the quotes and figures that matter against the original.

Long documents are where Claude earns its keep in a finance team. An annual report, a 90-page lease, a credit agreement, a policy manual: files you need answers from but don’t have time to read twice. Claude can read all of it at once. The trouble is that a smooth, confident summary of a 200-page file can be wrong in ways you’d never spot without reading it yourself, which defeats the purpose.

The fix is a method rather than a better prompt. Map the document first, pull facts out mechanically, ask questions that come back with page references, and summarize in two stages. Here’s how, with the prompts I use. For where long documents fit among Claude’s other uses in finance, see our guide to Claude for accounting firms and finance teams.

Product details in this guide were last checked against Anthropic’s official help pages in October 2026. Plans and features change often, so confirm on the vendor’s site before you buy or configure anything.

What Claude can take

Limit What Anthropic says What it means for you
Working memory (context window) 200,000 tokens on paid plans, roughly 500 pages of text. Several newer models handle 1 million A typical annual report fits in one conversation
PDF length Up to 1,000 pages Even very long filings can be uploaded whole
Charts and images in PDFs Read in PDFs of 100 pages or fewer. Over 100 pages, text only In a long report, figures inside charts or scanned tables won’t be read. Check those yourself or upload just those pages
File size 500 MB per file in a chat, 30 MB per file in a project Large scanned files may need to go in the chat, not the project
Files per chat Up to 20 Enough to compare several versions or agreements

The chart limit is the one to remember. Annual reports often put key numbers in graphics, and on a long file Claude simply won’t see them. Don’t assume it will tell you, so ask.

Step 1: Map before you dig

Before asking anything substantive, get the structure. You decide where to look next, not the model.

Produce the structure of this document: every heading and sub-heading
with its page number, in order. Do not summarize, interpret or comment.
If any pages are images or charts you could not read, list those page
numbers at the end.

The last sentence matters. It turns a silent gap into a list you can check.

Claude listing the headings and sub-headings of an annual report in order with a page number for each, followed by a list of pages it could not read
Our test with a public annual report. The map gives you the whole document’s structure with page numbers before you ask anything else.

Step 2: Pull facts out mechanically

Extraction is the job Claude does most reliably. Bring the facts out with their locations, then do the checking yourself.

List every monetary amount, date and deadline in sections [X] to [Y],
with the page number and the sentence it appears in.
Do not total, check or interpret them.
List every sentence in this document containing "must", "shall",
"will", "is required to" or "agrees to". Give the page, the party bound,
and the exact wording.

The second one is a mechanical sweep for obligations, and it’s more thorough than careful reading of a long file usually is.

Step 3: Ask questions that come back with page references

Using only this document, answer: [QUESTION].
After each sentence of your answer, give the page it came from.
If the document does not answer it, say so instead of inferring.

An answer you can check in ten seconds is worth five you’d have to re-read the document to verify. Per-sentence page references are what make that possible.

A page reference isn’t proof. It’s a faster way to check the proof yourself. Open the page.

Claude answering a question about revenue and operating income in an annual report, with a page number after each sentence
Same report, one question. Every sentence carries a page number, so checking the answer takes seconds instead of a re-read.

Be careful with “it doesn’t say”

“There’s no clause about X” is a much bigger claim than finding one, and less reliable. When Claude says something isn’t there, test it:

You said this document does not address [TOPIC]. List every term and
phrase you searched for, and quote the three passages that came closest.

Then run your own search in the PDF for the same words. If it’s important, absence needs two checks.

Step 4: Summarize in two stages, not one

Asking for a summary of a 200-page file in one go produces something that reads beautifully and is unevenly wrong. Some sections come through faithfully, others get squeezed into something the document never quite said, and nothing in the output tells you which is which.

Do it in two stages instead:

Summarize each top-level section in no more than [N] words, in document
order. Keep every figure exactly as written. Do not merge or reorder sections.
From these section summaries, write a [N]-word briefing for someone
who will not read the document: what it is, what it commits us to,
what to watch, and what is unclear. Cite sections throughout.

Mistakes stay inside one section, where they’re easy to find, instead of spreading through the whole briefing.

Diagram comparing a one-pass summary of a 200-page document, where faithful, compressed and wrong lines are mixed together, with a two-stage approach where section summaries are checked before a briefing is written
Same document, same model. Summarizing section by section keeps a mistake inside one section where you can find it.

Step 5: Check before you rely on it

  1. Spot-check the quotes. Open the PDF at three cited pages and confirm the wording matches exactly.
  2. Check the figures that matter against the source, not against Claude’s summary.
  3. Look at the pages it couldn’t read. Charts and scanned tables in long files are yours to check.
  4. Start a new conversation for a new question. In a long thread, an early misreading can carry through every later answer.
  5. Write down what you didn’t check. One line in the file: which document, which version, what was reviewed by hand, what wasn’t.

Our free AI output verification checklist turns these checks into a list you can tick off for any AI output.

Tips by document type

Document Ask Claude for Watch out for
Annual reports and 10-Ks The structure map, every figure on one topic with page references, what management says drove a change Key numbers inside charts and graphics, which aren’t read in reports over 100 pages
Leases Commencement and end dates, renewal and termination options, rent escalations, who pays for what Dates written as “within 30 days of” something else. Ask for every relative deadline separately
Credit and loan agreements Every covenant with its test date, every reporting obligation, the definitions behind each covenant Definitions. A covenant means whatever the definitions section says it means, so always pull both
Policy manuals Everything on one topic across all sections, and any two passages that seem to conflict Out-of-date versions. Confirm the version date before you rely on any answer

Before you upload the document

Long documents are often the sensitive ones: agreements under NDA, internal policies, unreleased reports. Run the five-question check in our guide to uploading client files to Claude first, and use your firm’s business account rather than a personal one.

Sources: official pages, checked October 2026

Frequently asked questions

Anthropic says paid plans have a 200,000-token context window, roughly 500 pages of text, and several newer models handle 1 million tokens. PDFs can be up to 1,000 pages, and up to 20 files can be uploaded per chat.

In PDFs of 100 pages or fewer, Claude analyzes both text and visual elements. In PDFs over 100 pages it processes text only, so numbers inside charts, graphics or scanned tables aren't read. Upload just those pages separately or check them by hand.

Ask for it explicitly: tell Claude to answer using only the document and to give the page number after each sentence, and to say so if the document doesn't answer the question. Then spot-check a few of the cited pages yourself.

A one-pass summary of a long file can render some sections faithfully and compress others into things the document never said, with nothing to show which is which. Summarizing each section first and then writing the briefing from those summaries keeps any mistake inside one section where it's easy to find.

It can try, but saying something isn't there is less reliable than finding something that is. Ask which terms it searched for and which passages came closest, then search the PDF yourself for the same words.

Testing methodology

Last tested Oct 9, 2026
Versions tested Claude Pro (Claude desktop app, Opus 5.5) File, page and context limits checked against Anthropic's help pages, October 2026

We uploaded a public company annual report, a PDF of more than 100 pages, to a Claude Pro account and ran two of the prompts in this guide in the same conversation.

  • The structure map listed the headings and sub-headings in order with their printed page numbers, and noted that the cover and back cover are unnumbered.
  • The page-referenced question returned total revenue and operating income for the latest fiscal year and the drivers the report gives, with a page reference after every sentence.

File, page and context-window limits are documentation analysis, checked against Anthropic's help pages in October 2026.

What we did not test here: scanned, image-only PDFs, and the two-stage summary on this report.

Common errors and what they mean

If you're hitting an error or unexpected behavior, here's what each one means and the fastest path to resolving it.

Error Why it happens How to fix it
The summary misses numbers shown in chartsPDFs over 100 pages are processed as text onlyAsk for unreadable pages to be listed and check those figures by hand
A confident summary is wrong in placesA one-shot summary of a long file compresses sections unevenlySummarize section by section, then summarize the summaries
"The document doesn't mention X" turns out to be wrongAbsence is harder to establish than presenceAsk which terms it searched for, then search the PDF yourself
A page reference points to the wrong pagePrinted page numbers differ from the PDF's own page countAsk which numbering it used and open the cited page to check
Later answers repeat an early misreadingA long conversation carries earlier mistakes forwardStart a new conversation for each new question
FREE · 7 PAGES

Long-Document Analysis: Working Through 200-Page Files

Sent straight to your inbox. No account, no upsell.

No spam. One reply unsubscribes you.