Claude PDF Skill: Native Reading vs. the Skill

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 that teaches Claude to manipulate PDFs as files — merge, split, fill forms, rotate pages, add watermarks, rebuild a document from scratch. Confusing the two leads to two failure modes: people who never touch the skill and manually stitch together documents Claude could assemble in one prompt, and people who install the skill 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.

Native PDF reading already does more than most people assume

Upload a PDF to Claude — in the chat app, Claude Code, or the API — and Claude reads each page as both an image and an extracted text layer at once. That means it sees layout, tables, seals, handwriting, and charts 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 current limits worth knowing: PDFs up to 32MB and up to 100 pages per document work natively, per Anthropic's PDF support documentation. For anything longer, split it into parts or summarize section by section — Claude's large context window means you can usually paste several chunks into one conversation and still get a coherent answer across the whole document.

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. Per Anthropic's published skill, the core capabilities are: extracting text and tables programmatically, merging and splitting documents, rotating and reordering 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.

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 — and adding the skill does nothing but slow the first response down. See our roundup of the best Claude skills for where this one ranks against everything else worth installing.

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.

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 installing 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 add the pdf skill to the conversation or Project. 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.

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. 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. 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.

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. 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. 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, Claude Code, or through the API — there is no toggle to enable and nothing to install. The pdf skill is a separate, optional addition you only need for file manipulation tasks like merging, splitting, or form-filling, not for reading and analysis.

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. Once quality drops — skewed pages, faint print, handwriting mixed with print — 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?

Functionally, yes, for the operations it covers. It cannot do everything a full desktop PDF editor does — there is no visual layout designer — but for merging, splitting, rotating, watermarking, encrypting, filling form fields, and rebuilding a document from extracted content, it replaces the editor entirely inside a conversation, with no separate software to open.

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.