TL;DR
Drop a file in and ask “summarize” and Claude gives you the same hedged, safe-sounding output any model produces. The difference between a usable answer and a dangerous one is whether Claude grounded in the source. The workflow: rename the file, inventory before interpreting, drill section by section, and demand a page or section citation for every claim. As of June 2026 Claude.ai chat takes 30MB per file and up to 20 files per conversation; Opus 4.7 and Sonnet 4.6 carry a 1M-token context window (~3,000 pages). The rest of this guide is the prompt sequence and the verification step most people skip.
Who this is for
Anyone uploading PDFs, contracts, research papers, transcripts, code, or spreadsheets into Claude and expecting answers they can act on — legal, finance, research, M&A diligence, journalism. If a wrong number or a hallucinated clause has real cost, this is the workflow.
When to reach for it (and when not to)
Reach for it on long files (50+ pages), multi-file comparisons, anything you need to cite by passage, files you’ll reference repeatedly, or any case where a fabricated answer is expensive. Claude is genuinely stronger here than ChatGPT for grounded document Q&A — but only if you upload well.
Skip it for files with sensitive PII or non-disclosable info, files past the upload limits, quick questions a search would answer, and binary formats with no text layer (raw images, encrypted PDFs — Claude rejects password-protected PDFs outright).
Claude file limits as of June 2026
These are the numbers that actually constrain your workflow. The chat app and the developer API have different ceilings, which trips people up.
| Surface | Max per file | Files at once | Page / context ceiling |
|---|---|---|---|
| Claude.ai chat (Free/Pro/Max) | 30 MB | 20 per conversation | Bounded by model context (1M tokens on Opus 4.7 / Sonnet 4.6) |
| Claude Projects | 30 MB | Effectively unlimited in project knowledge | Retrieves relevant chunks; full set won’t fit context at once |
| Anthropic API (inline PDF) | 32 MB request | — | 600 pages (100 on 200k-context models) |
| Files API (developers) | 500 MB | 100 GB per workspace | Reference by file_id to keep requests small |
Supported document formats in chat: PDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB, JSON, and XLSX (XLSX needs code execution enabled). Images: JPEG, PNG, GIF, WebP up to 8000×8000 pixels. Source: Anthropic’s Upload files to Claude help page and the PDF support API docs.
One detail that explains most “Claude misread my chart” complaints: every PDF page is processed as both extracted text and a rendered image, costing roughly 1,500–3,000 tokens per page. Dense, small-font, or graphics-heavy pages can exhaust the context window before you hit any page count. If a file feels like it’s being skimmed, it probably is — split it.
Before you start
- Rename files for clarity:
contract-vendor-a-2026.pdf, notFinal_v3 (1).pdf. Claude quotes filenames in answers, so readability pays off. - Strip non-essential pages. A 600-page report with a 580-page appendix grounds worse than the 20 pages you actually need.
- Decide a citation format up front — page numbers, section headers, or both — and state it in your first prompt so Claude stays consistent.
- Keep an independent copy open in another window to spot-check claims. Trust nothing on the first pass.
The workflow, step by step
- Rename before uploading. Give the file a name Claude can reference cleanly.
- Upload a set in one message. For multi-file work, attach all files together so Claude treats them as a group and you can ask comparative questions immediately.
- Inventory first, interpret never (yet). First message after upload:
Describe what each file is, its sections, and rough size. Do not summarize content yet.This forces a structural pass before any interpretation. - Ask for the table of contents, then drill. For long PDFs, get the TOC, then work one section at a time. Broad questions across a full document degrade silently in the middle — a well-documented “lost in the middle” effect even at 1M tokens.
- Demand a citation for every claim. Phrase:
Quote the relevant passage and the page or section header.Refuse answers that paraphrase without a source. - Force structure on tables and numbers. Ask for CSV or Markdown output, not prose — it’s easier to verify. Recompute one row by hand on numbers that matter.
- Promote recurring files to a Project. When you’ll keep working with a file, put it in a Project so you don’t re-upload each session. See Claude Projects and the advanced Projects workflow.
Prompts that actually ground Claude
Save these as a reusable set. They map directly to the steps above.
Inventory (run first, every time):
Before answering anything: describe each uploaded file — its title,
type, page or row count, and section headers. Do not summarize or
interpret content yet.Section drill with mandatory citation:
Working only from Section 4 of [filename]: list every quantitative
claim. For each, quote the exact passage and give the page number or
section header. If a number is not in the source, say so — do not infer.Audit-grade extraction (scriptable):
Output findings as a JSON array. Each item must have: claim, file,
page, quote, confidence (low | med | high). Include only claims you can
quote verbatim from the source.Two-contract redline:
File 1 = current contract. File 2 = proposed redline. List every
changed clause. For each, quote the old text with its page and the new
text with its page. Flag any clause removed entirely.Verify it before you trust it
This is the step that separates real work from “looked right.” On a document that matters, do not skip it.
- Every numeric claim or quote carries a page or section reference. No exceptions.
- Manually check at least 3 references against the original. Pick the most surprising claims — that’s where hallucinations hide.
- Confirm what the file does not contain. Ask
What sections are NOT in this file?as a sanity check against fabricated structure. - In multi-file work, every answer names its source file. If Claude blends files, push back with explicit filename anchors.
A quick calibration drill: pick a 30–50 page document you’ve already read, run the inventory and citation prompts on a fresh chat, then count how many of Claude’s cited pages actually contain the claim. If you’re below 90%, tighten the citation requirement and rerun before trusting it on anything new.
Common mistakes
- Asking for a “summary” before knowing the file’s structure. You get safe-sounding nothing.
- Mixing 10 unrelated files in one upload. Claude blends sources in its answers.
- Trusting unsourced claims. Always re-ask for the exact passage and page.
- Treating OCR as perfect. Heavily scanned PDFs garble passages — ask Claude to output a suspect page as raw text and eyeball it.
- Loading the full document once and firing one broad question. Middle-of-document recall degrades quietly.
- Letting old versions pile up in a Project. Stale files blend with current ones and corrupt answers; prune them.
FAQ
- What’s the max file size? As of June 2026, 30MB per file and 20 files per conversation in Claude.ai chat. Developers using the Files API get 500MB per file. Past the limit, split the file or extract the relevant sections first.
- How many pages can a PDF be? In chat, you’re bounded by the model’s context, not a hard page cap. Via the API, the limit is 600 pages per request (100 on 200k-token models), within a 32MB request body. Each page costs ~1,500–3,000 tokens.
- Do uploaded files train Claude? By default Anthropic does not train its models on consumer chat or uploaded content. Check your specific plan and any Team/Enterprise data settings to confirm.
- Can Claude handle Excel with formulas? It reads values, and with code execution enabled it can compute on XLSX. It does not audit formula logic the way a spreadsheet tool does — for that, export to CSV and ask explicitly.
- What about scanned PDFs with bad OCR? Quality drops on dense scans. Because each page is also rendered as an image, Claude often recovers text the OCR mangled, but verify: ask it to output the suspect page as raw text and spot-check.
- How do I compare two contracts cleanly? Upload both at once and give each a role: “File 1 = current contract, File 2 = proposed redline; list every changed clause with quotes and page numbers from each.” Use the two-contract prompt above.
How to reuse this workflow
Save the prompt set as named templates (“Standard contract review,” “Research-paper extraction”). For recurring file types, build a Project with custom instructions like Always cite page numbers; refuse to answer without a source. Keep a running list of past hallucinations — failure modes recur, and naming them helps you catch them faster. Every quarter, re-test with a fresh file to confirm Claude’s behavior hasn’t drifted after a model update.