AItomation Academy
← All posts
Claude Workflows

Claude PDF Skill: Native Reading vs. the Skill

Marko Sudar·

Get The Claude Starter System

20 prompts, skills, and guides that get finished work out of Claude. Free PDF.

Claude can already read a PDF the moment you drop it into a chat — text, tables, charts, even scanned pages — with no setup at all. The Claude PDF skill is a different, narrower thing: a packaged set of instructions and scripts that teaches Claude to manipulate PDFs as files — merge, split, fill forms, rotate pages, add watermarks, rebuild a document from scratch. It is one of Anthropic's built-in skills, so as of October 2026 there is nothing to download or install on claude.ai: with code execution switched on, Claude calls it by itself when a request needs it. Confusing reading with the skill leads to two failure modes: people who never ask Claude to touch the file and manually stitch together documents Claude could assemble in one prompt, and people who go hunting for a PDF skill to add for a job — like "summarize this 12-page report" — that native reading already handles better and faster.

This guide draws the line clearly, then gives you two full recipes that sit right on either side of it: contract review (native reading does the heavy lifting) and multi-file report extraction into a clean spreadsheet-ready table (where the skill earns its place). If you are new to the skills system generally, start with what Claude Skills are before this one — this article assumes you know the basic concept.

Can Claude read PDFs without the PDF skill? Yes

Yes. Upload a PDF in the Claude app or send one through the API, and Claude reads each page as both an image and an extracted text layer at once. That means it sees tables, charts, diagrams, and images the same way a person would, not just the text a copy-paste would grab. No skill, no setup, no extra step.

In practice this covers the majority of what people actually need from a PDF: summarizing it, answering questions about it, pulling specific figures or clauses, comparing it against another document, rewriting a section, or flagging what is missing. If your job is "understand this document and tell me something about it," native reading is the whole answer — you do not need a skill for that, and installing one adds a step for no benefit.

The limits worth knowing, as of October 2026: the Claude apps accept up to 20 files per chat, up to 500MB per file, and PDFs up to 1,000 pages, per Anthropic's help center. Claude analyzes charts, images, and other visual elements only in PDFs of 100 pages or fewer; from 101 to 1,000 pages it reads the text alone. Files added to a Project's knowledge are capped at 30MB each. Through the API, Anthropic's PDF support documentation sets 32MB per request and 600 pages per request (100 pages when the request's context window is under 1M tokens), and it lists password-protected or encrypted PDFs as unsupported.

Where you uploadFile sizePagesWhat Claude sees
Claude app chat500MB per file, 20 files per chatUp to 1,000Text and visuals up to 100 pages; text only from 101 to 1,000
Project knowledge30MB per fileContext window (up to 10x more with RAG on paid plans)Reusable across every chat in the Project
Claude API32MB per request600 (100 under a 1M-token context)Text plus an image of every page

Limits as published by Anthropic, October 2026.

The 100-page line is the one that catches people. A 250-page scanned report uploaded to a chat lands in the text-only band, and a scan has no text layer, so Claude can come back with almost nothing to work from. There are two fixes: split the file into parts of 100 pages or fewer so every page gets read as an image, or have Claude run OCR first (a pdf skill job, covered below) so the pages carry real text. For long PDFs that already have a text layer you need neither, though dense documents can still fill the context window before they reach the page cap. When that happens, split by section and ask for a summary of each part in the same conversation.

Where the pdf skill actually earns its keep

The pdf skill is for when the output needs to be a PDF, or the input needs surgery before Claude can read it usefully. That is a meaningfully different job from reading and reasoning about content. The skill works by running code in Claude's sandbox, using Python tools such as pypdf, pdfplumber, and reportlab plus Tesseract for OCR, and it hands you a new file to download. Per Anthropic's published skill, the core capabilities are: extracting text and tables programmatically, merging and splitting documents, rotating pages, filling in form fields, adding watermarks, encrypting or decrypting files, extracting embedded images, running OCR on scanned pages that have no text layer at all, and assembling brand-new PDFs from scratch. One exception worth knowing: if the PDF you need is really a presentation, Claude Slides has its own PowerPoint and PDF downloads for finished decks, so you do not need to ask for the pdf skill there.

The pattern to watch for: if your request has a verb like "merge," "split," "fill," "rotate," "watermark," "OCR," or "rebuild," you want the skill. If the verb is "summarize," "explain," "compare," "find," or "flag," you want native reading. You do not need to switch the skill off for those jobs either: until a request calls for it, Claude keeps only the skill's name and short description in view, about 100 tokens per skill according to Anthropic's Agent Skills documentation, so having it available costs next to nothing. See our roundup of the best Claude skills for where this one ranks against everything else worth installing.

One quirk to expect: the version of the skill Anthropic publishes describes itself broadly, telling Claude to use it "whenever the user wants to do anything with PDF files." So Claude sometimes writes code to pull text out of a file it could simply read. That is harmless but slower. If it happens on a reading job, add "read the PDF directly, no code" to your prompt and ask again.

Native Reading vs. the PDF Skill

Use native reading (no setup)

  • Summarize or explain a PDF
  • Answer questions about its contents
  • Pull specific numbers, clauses, or dates
  • Compare two or more documents
  • Flag risks, gaps, or inconsistencies
  • Read tables, charts, and clean scans (visuals up to 100 pages in chat)

Use the pdf skill

  • Merge or split multiple PDFs
  • Rotate, reorder, or delete pages
  • Fill in PDF form fields
  • Add a watermark or encrypt a file
  • OCR a scanned document with no text layer
  • Remove a password you know so Claude can read the file
  • Assemble a brand-new PDF from scratch

How do you turn on the Claude PDF skill?

On claude.ai you do not install the Claude PDF skill: it is built in, and Claude uses it automatically once code execution and file creation are switched on. The setup differs a little depending on where you use Claude.

In the Claude apps, skills are available on every plan, Free included, and they need code execution. On Free, Pro, and Max, go to Settings > Capabilities and turn on "Code execution and file creation." On Team and Enterprise, an owner turns on "Cloud code execution and file creation" and "Skills" under Organization settings > Plugins & skills (Team has skills on by default). After that, Claude decides when the pdf skill is needed based on your request, and you can see or toggle individual skills under Customize > Skills, per Anthropic's guide to using skills in Claude.

In Claude Code, which needs a paid plan, the same built-in skill comes along when you sign in with your Claude account. The Claude Code skills documentation says Anthropic's pdf and xlsx skills always sync to signed-in sessions: run /skills to see pdf listed under "claude.ai sync," or call it directly with /pdf. Sessions that sign in with an API key or run through a cloud provider such as Amazon Bedrock do not sync it. If you want your own variant, such as a skill with your company's form-filling rules baked in, our guide to Claude Code skills covers where skill folders live.

Developers calling the Claude API get the same skill by passing the skill ID "pdf" alongside the code execution tool. Everyone else can ignore that route: the Claude app covers every recipe in this guide.

Recipe 1: Contract review with native reading

Contract review is a native-reading job almost every time — you are not editing the PDF, you are interrogating it. Upload the contract, then use a prompt that forces Claude to structure its findings by risk rather than just narrate the document back to you in order.

You are reviewing a contract for a non-lawyer business owner. I have
attached the PDF. Read the entire document before answering.

CONTEXT: [what this contract is for — e.g. "vendor agreement for
a marketing agency we are hiring for 6 months"]
MY ROLE: [buyer / seller / employer / employee / other]
MY PRIORITIES: [e.g. "exit flexibility, IP ownership, liability cap"]

Produce:

1. PLAIN-ENGLISH SUMMARY (5-6 sentences)
   What this contract commits each party to, in the order that
   matters, not the order it appears in the document.

2. RISK FLAGS (ranked, most concerning first)
   For each: the exact clause or section reference, what it says,
   why it matters for someone in MY ROLE, and how unusual or
   aggressive it is compared to a standard version of this contract.

3. MISSING OR VAGUE TERMS
   Anything a contract like this would normally specify but this one
   does not, or specifies ambiguously.

4. NUMBERS AND DATES TABLE
   Every dollar amount, percentage, deadline, and term length in the
   document, with its section reference, in one table.

5. QUESTIONS TO ASK BEFORE SIGNING
   3-5 specific questions to raise with the other party, phrased so
   a non-lawyer can ask them without sounding adversarial.

Quote the exact contract language for anything in sections 2 and 3 —
do not paraphrase clauses you are flagging as risky. If a section is
hard to read because of scan quality or dense formatting, say so
rather than guessing at its content.

Two things make this work well. First, quoting exact language back forces Claude to point at the actual clause instead of summarizing loosely — you can verify every flag in seconds. Second, splitting risks from missing terms surfaces the omissions a first read usually misses, which is often where the real exposure sits. This is analysis and review, not editing, so no skill is involved — it is native reading doing the entire job.

The Claude Vault — 550+ prompts & skills

Skip writing the prompts yourself

  • 500+ copy-paste prompts & workflows
  • 35 installable Cowork skills
  • Role packs for 6 professions
  • New prompts & skills every week, forever

Recipe 2: Report extraction across many PDFs

This is where the line between native reading and the pdf skill gets interesting. If you are extracting data from one report Claude can read natively — no skill needed, just a structured extraction prompt. The skill becomes worth calling on when the job is repeatable across a batch of scanned reports with no text layer, or when the end product itself needs to be a PDF (a combined packet, a cleaned-up version with pages reordered, or a watermarked copy for distribution).

Start with the extraction prompt, which works with native reading alone for a handful of readable PDFs:

I am attaching [N] reports as PDFs. Extract the following into one
consolidated table, one row per document:

FIELDS TO EXTRACT:
- Document name / source
- Report date or period covered
- [Field 1 — e.g. "total revenue"]
- [Field 2 — e.g. "headcount"]
- [Field 3 — e.g. "top risk mentioned"]
- Page/section reference for each extracted figure

RULES:
- If a field is not present in a document, write "not found" — do
  not estimate or infer a number that is not explicitly stated.
- If a number appears more than once with different values (e.g.
  a preliminary figure and a restated figure), report both and flag
  the discrepancy.
- Where a document is a scanned image with no reliable text layer,
  say so explicitly instead of guessing at unclear numbers.

Output as a markdown table I can paste into a spreadsheet, plus a
short note below it flagging any document where extraction was
low-confidence.

If that last rule keeps triggering — several of your reports are scans with no text layer, or you need the final consolidated output saved back out as an actual PDF packet — that is your signal to bring in the pdf skill. You do not attach it to the conversation or Project; you ask for the operation, and Claude loads the skill when it sees the request. Ask Claude to OCR the scanned pages first, then rerun the same extraction prompt against the OCR'd text; the skill handles the conversion, the prompt above still does the extraction work unchanged. Once the table checks out, one more line, such as "merge the source PDFs into one packet in the order of the table and add a cover page listing them," gets you the PDF back out. After extraction, you can paste the table into a workbook and analyze it in Excel with Claude.

Turning the recipe into a repeatable workflow

For either recipe, the highest-leverage move is not the individual prompt — it is putting it in a Project so you are not rewriting the setup every time. Create one Project for "contract review" and one for "report extraction," save the prompt above as a Project instruction, and every new document just gets dropped in with the specific context filled out. Projects are on every plan, though Free is capped at five, and on paid plans Claude switches a Project to retrieval (RAG) mode as its knowledge nears the context limit, expanding capacity up to 10x. Once a prompt is stable, you can go a step further and turn it into a skill of your own that fires in any chat, no Project needed; our guide on how to build a Claude skill walks through the SKILL.md format. If your documents live in Obsidian rather than loose PDF files, our Claude skill for Obsidian guide covers wiring that vault in directly instead of exporting files by hand each time.

The same before-you-automate discipline applies here that applies to any document workflow: run the prompt manually on 3-5 real documents first and check the output against the source before you trust it on a stack of 40. A wrong number in a contract-risk flag or a misread revenue figure is expensive precisely because it looks confident. A stack that size also runs past the app's 20 files per chat, so either work in batches or, on a paid plan, connect the folder in the desktop app and hand the whole job to Claude as a longer task (the agentic work Anthropic launched as Cowork, which is rolling out as part of Claude itself). If you are also working from meeting recordings or interview transcripts alongside these documents, our comparison of ChatGPT vs. Claude for summarizing transcripts is a useful companion piece — same verification logic, different input format.

Marko Sudar — founder of AItomation Academy

Want workflows like this built and updated weekly?

  • 1,300+ non-technical professionals learning together
  • Claude, Claude Cowork & Claude Code training
  • The full 550+ prompt & skill Vault included with Premium
  • A weekly live call to get unstuck

The short version

Default to native reading — it is faster, needs no setup, and covers summarizing, answering questions, comparing, and flagging risk across contracts and reports alike, with charts and images read on PDFs up to 100 pages in chat. Reach for the pdf skill only when the job involves manipulating the file itself: merging, splitting, form-filling, watermarking, encrypting, or OCR-ing scans with no text layer. It is built into Claude, so the only setup is turning on code execution. Most people who think they need the skill actually need a better extraction prompt, and most people who never touch it are missing an easy win the first time they have to merge twelve scanned invoices into one clean document. For the full picture of what else is worth installing alongside it, revisit what Claude Skills are and the best Claude skills to round out your setup.

Frequently asked questions

Do I need to install anything to have Claude read a PDF?

No. Native PDF reading works the moment you attach a file in Claude.ai or send one through the API — there is no toggle to enable and nothing to install. The pdf skill is not an install either: it is built into Claude, needs only code execution and file creation turned on, and handles file manipulation tasks like merging, splitting, or form-filling, not reading and analysis. Claude Code can open PDFs that sit in your project folder too, though PDFs longer than 10 pages need poppler installed on your machine.

Can Claude read a scanned PDF with no selectable text?

Sometimes, but not reliably. Native reading treats each page as an image, so it can often make sense of a clean scan on its own, though in the Claude apps that image reading only applies to PDFs of 100 pages or fewer. Once quality drops — skewed pages, faint print, handwriting mixed with print — or the scan runs past 100 pages, accuracy suffers, and that is exactly the case where running the pdf skill's OCR step first, then handing the cleaned-up text back to Claude, produces a more reliable result than either step alone.

Is the pdf skill the same as a PDF editor?

Not quite. It covers merging, splitting, rotating, watermarking, encrypting, filling form fields, and rebuilding a document from extracted content by running code on the file and handing you a new PDF to download. There is no visual layout designer and no click-to-edit text, so for changing one sentence in place a desktop PDF editor is still the better tool; for batch jobs like those above, the skill saves you opening separate software at all.

Will Claude ever get the extraction wrong on a contract or report?

Yes, occasionally, especially with dense financial tables, footnotes that modify a headline number, or clauses that reference definitions elsewhere in the document. That is why both recipes above ask Claude to quote exact language and flag low-confidence extractions rather than smoothing over uncertainty. Treat the output as a strong first pass that still needs a human check on anything that will inform a decision, not a final answer to forward untouched.

Does the Claude PDF skill work on the free plan?

Yes. As of October 2026, skills, file creation, and code execution are all included on Claude's Free plan, so the pdf skill can merge, split, fill, and build PDFs there. Free has tighter usage limits, runs on Sonnet and Haiku only, and does not include Claude Code; our Claude pricing guide breaks down what each paid plan adds.

Join the Academy

1,300+ professionals, the courses, the full Vault, and a weekly live call.

Join the Academy

Grab the playbook

The Claude Starter System — free PDF to your inbox.

Claudeclaude pdf skillclaude pdfpdf skillclaude skillscontract reviewdocument extractionpdf workflowsclaude for business