TL;DR
Upload the document, ask for a structural outline first, then drill into one section at a time with page-cited quotes, check what is missing, verify 3-5 quotes by hand, and only then ask for the executive summary. On a 200-page document the full loop runs 30-60 minutes and produces a summary you can defend in a meeting. The single biggest mistake is “summarize this 200-page PDF in 5 bullets” on the first turn: you get a confident, plausible answer that quietly skips the section that mattered.
What this covers
A repeatable workflow for processing long documents (roughly 100-500 pages) in Claude without splitting, mis-paraphrasing, or hallucinating sections that do not exist. Claude’s long context is genuinely strong on documents like these, but only with structured prompting. This guide gives you the structure and the exact prompts.
Key tools and concepts:
- Claude (Opus 4.7 / Sonnet 4.6): Anthropic’s models. On the claude.ai chat interface, paid plans run a 500K-token context window with these models; the full 1M-token window is available through Claude Code and the API (as of June 2026). Sonnet 4.6 is the faster workhorse; Opus 4.7 is the more careful reader for high-stakes documents.
- Structural outline pass: Asking for the document’s structure before any content interpretation. This is the single most important step for grounded answers.
- Per-section drill-down: Asking about one section at a time with page-cited claims, instead of one broad question across the whole file. This counters the “lost-in-the-middle” recall drop that still affects every long-context model in 2026.
Who this is for
Researchers, analysts, lawyers, students, and anyone who reads long PDFs for a living. Especially valuable if you have ever been burned by an LLM “summary” that left out the part you needed most.
When to reach for it
Reading reports, contracts, papers, depositions, technical specs, regulatory filings, or anything past 50 pages. Also useful for multi-document comparisons when each document is itself long.
The limits you are actually working inside (June 2026)
Know the box before you fill it. These are the claude.ai consumer limits as of June 2026; the API limits are higher.
| What | claude.ai chat (Pro/Max) | Claude Code / API |
|---|---|---|
| Context window (Sonnet 4.6 / Opus 4.7) | 500K tokens | 1M tokens |
| Max file size per upload | 30 MB | 32 MB request / Files API for larger |
| Files per conversation | 20 | n/a (Projects: effectively unlimited storage) |
| Pages per PDF (full vision analysis) | best under ~100 pages | up to 600 pages/request |
| Tokens per PDF page | ~1,500-3,000 (each page is parsed as text and an image) | same |
Two practical consequences. First, a dense 100-page PDF can consume 150K-300K tokens once every page is rendered as an image, so a single big contract plus a few reference files can fill even the 500K chat window faster than the page count suggests. Second, Projects do not load the whole knowledge base into context — Claude retrieves the relevant slices at query time, which is why page-anchored questions beat “summarize everything.” (Source: Anthropic PDF support docs.)
Before you start
- Confirm the document has a real text layer (try selecting text in a PDF reader). Heavy scans rely on OCR, which adds noise.
- Have an external copy open in another window to spot-check claims. Trust nothing on the first pass.
- Decide your output target: an executive summary, a section-by-section table, an extracted quote list, or a compliance gap analysis. Different targets need different prompting.
- Carve out 30-60 minutes. Long-doc work is not a 90-second task; do not squeeze it.
Step by step
- Upload and name it. Rename the file to something memorable (
acme-vendor-msa-2026.pdf) so Claude can reference it by name in later messages. For documents you will revisit, put them in a Project rather than a one-off chat. - Ask for a structural outline first — no content yet:
List every top-level section and subsection of this document with the page range each covers. Do not summarize content.This forces inventory before interpretation and gives you a map to drill against. - Drill into one section at a time, with page references:
In Section 4 (pages 23-41), list every numbered obligation with its exact phrasing and page number.Repeat only for the sections that matter to your goal. - Check the negative space. Ask what is not there:
What sections would you expect in a vendor MSA that this document does NOT contain? Be specific.This catches absent terms you might assume were present. - Verify by hand. Check at least 3 quoted passages and 2 numbers against the source. If any are wrong, push back before continuing — accuracy degrades after the first uncorrected hallucination.
- Ask for the executive summary last, only after section-level details are confirmed:
Now write a one-page executive summary. Cite the section and page for every claim.
A worked example, for a 60-page services agreement: step 2 returns an 11-section outline; you drill steps 3-4 on Sections 4 (Fees), 7 (Termination), and 9 (Liability); the negative-space check surfaces that there is no data-breach notification clause; you verify the liability cap quote and the payment-terms number by hand; then you request the summary. Total time: about 25 minutes.
Why one section at a time beats one big question
Long-context models still lose information in the middle of the input. As of 2026, frontier models including Opus 4.7 score around 90% on single-needle retrieval at 1M tokens, but multi-fact (“multi-needle”) recall is meaningfully lower, and facts sitting at 30-70% positional depth show a measurable 5-15 point retrieval drop versus facts near the start or end. The practical effective context for reliable multi-fact work sits roughly in the 200K-400K band, even though the window is 500K or 1M. Pinning each question to a section and page range keeps your target near the model’s attention and sidesteps the soft middle. See Claude long-context unstable for the failure pattern in detail.
Quality check
- Every claim has a page or section reference. No exceptions on documents that matter.
- You have verified at least 3 references by hand. Pick the most consequential claims.
- Claude can name what is not in the document (a sanity check against hallucinated coverage).
- The executive summary’s claims trace back to specific sections, not generic phrases like “the document discusses.”
- For numeric claims, recompute one totals row by hand. LLMs are most confidently wrong on arithmetic.
How to reuse this workflow
- Save the prompt sequence as a template (“Long-PDF intake v3”).
- For recurring document types (contracts, papers, filings), customize the template with type-specific questions and a “common omissions” list — things you wish you had asked. Add to it every session.
- For repeated review of the same documents, keep them in a Project so you skip re-uploading; Claude retrieves the relevant pages per query.
- Re-test with a known-good document each quarter; model behavior and limits evolve.
FAQ
- How long is “long” for Claude?: As of June 2026, claude.ai chat gives paid plans a 500K-token window with Sonnet 4.6 / Opus 4.7; Claude Code and the API reach 1M tokens. 500K is roughly 1,000-1,500 pages of plain prose, but a dense PDF parsed as page images fills it faster. Past the practical limit, split by section, not by raw page count.
- Does long context degrade in the middle?: Yes. Single-needle recall is ~90% at 1M tokens, but multi-fact recall drops and mid-document facts lose 5-15 points (Claude long-context unstable). Per-section drill-down mitigates it.
- Should I ask one section at a time, or everything at once?: One at a time for accuracy. All-at-once only when you already know the document and just need a refresh.
- What about scanned PDFs?: Each page is processed as text and an image, so clean scans usually work, but heavy or skewed scans lose accuracy. Spot-check OCR and ask Claude to output suspect pages as raw text.
- Can Claude compare two long documents at once?: Yes, but expect more drift. Give each a clear role in the prompt (“File 1 = old contract, File 2 = new redline”) and run section-by-section comparisons rather than one global diff.
- How do I keep a long document for future sessions?: Put it in a Project (see Claude Files). Projects store files persistently and retrieve relevant sections at query time.
Common mistakes
- “Summarize this 200-page doc in 5 bullets” — you get a confident, plausible, partially wrong summary.
- Not requesting citations, so every claim becomes equally trustworthy and equally suspect.
- Trusting numbers without spot-checking. LLMs are most confident exactly where they are most likely to be wrong.
- Asking one question of a 500K-token paste and assuming the answer covers the middle of the document — recall drops there unless you pin the question to a section (Claude long-context unstable).
- Treating Claude’s “I cannot find this” as definitive — sometimes it just missed; rephrase with section anchors.
- Stopping at the executive summary without verifying its citations against the section drill-downs.