Baonat
← All skills
Documents · ClaudeVerified

Author: Anthropic

Claude PDF skill — Read, merge, fill, OCR

Work inside real PDFs: extract tables, merge/split files, fill forms, watermark, and OCR scans — instead of pasting page dumps into chat and hoping layout survives.

Access: Claude paid tiers often ship PDF document skills built-in (or via the document-skills plugin). The skill artifact on GitHub is free to browse/clone; host Claude product plans may still apply for runtime use.

What it does
Loads Anthropic’s pdf Agent Skill so Claude follows document-native steps for create/read/edit/analyze on .pdf — including form fields and scanned pages when OCR is needed.
Best for
Ops, legal ops, finance, and support teams that live in invoices, contracts, form packs, and scanned packets.
How to install / use
Enable Anthropic’s pdf skill from anthropics/skills (or Claude document-skills plugin / built-in document skills on eligible paid plans). Attach the PDF + goal; ask for a page map before bulk edits.

Attach pdf before you ask Claude to merge, fill, or OCR — not after the first messy extract.

Copy this

  • Ask for a page map (page # → content type) before bulk edits
  • State output: single PDF vs split files vs extracted tables
  • For scans: require OCR pass + confidence notes on unreadable fields

When to use

Use Anthropic’s official pdf Agent Skill when the work product is a real PDF — extract, merge, split, fill forms, watermark, or OCR a scan — not a markdown paraphrase of what was on the page.

Good fits:

  • Invoice / statement packs that need table extraction into a clean grid.
  • Form packets (applications, onboarding) where fields must be filled and the file returned as PDF.
  • Merging exhibits or splitting a binder by section.
  • Scanned contracts or receipts that need an OCR pass before analysis.

Skip it for one-paragraph summaries of a short text PDF you can paste, or when the destination is Word/Sheets and you should use the matching document skill instead.

Access note

Claude paid tiers often include PDF document skills built-in, or via Anthropic’s document-skills plugin. The skill folder on GitHub is free to browse and clone as a reference artifact. Baonat lists it because host-subscription users are the audience — we do not sell a separate paid skill pack. Runtime still depends on your Claude plan or API surface.

Why

Bare chat turns PDFs into flat text and loses forms, page order, and “which page was that table on?” Anthropic’s public pdf skill (in anthropics/skills) keeps PDF operations first-class: treat the file as the deliverable, with scripts and checklists for merge/split/fill/OCR instead of hoping the model invents layout.

Baonat marks this verified. Artifact on GitHub is free; Claude product billing may still apply when you run it.

Install

  1. Get the skill from skills/pdf, or install Anthropic’s document-skills plugin / use Claude’s built-in document skills per Agent Skills overview.
  2. State the job type up front: extract · merge · split · fill form · watermark · OCR.
  3. Attach inputs with constraints: page ranges, field names, output naming, “do not invent values.”
  4. Page map first for multi-page jobs — confirm structure before bulk edits.
  5. Verify output in a real PDF reader (forms still fillable? page count correct?).
  6. Human review for legal/finance numbers before send.

Copy this

Kickoff prompt:

Attach the pdf skill. Job: [extract|merge|split|fill|OCR|watermark].
Files: [attach]. Constraints: [page ranges / field list / naming].
First response ONLY: page map (page → type: text|table|form|scan) + proposed steps.
Do not invent filled values. Flag unreadable OCR fields with confidence notes.
Deliver: [single PDF | split files | extracted tables]. End with a short verification checklist.

Concrete scenarios

  1. Month-end invoices. Extract line items from a PDF pack into a table; keep source page references.
  2. Onboarding forms. Fill known fields from a structured intake; leave unknowns blank and listed.
  3. Binder split. Split a merged PDF into section files named by TOC headings.

Pitfalls / when NOT to use

  • Blind OCR trust. Always flag low-confidence fields.
  • 80-page one-shot rewrite. Chunk by section or page range.
  • Wrong format. If the client needs .docx or .xlsx, use those skills instead.

FAQ

Q: Official Anthropic skill?
A: Yes — pdf under anthropics/skills, plus Claude Agent Skills docs.

Q: Is the GitHub skill free?
A: The artifact is free to obtain from GitHub. Paid Claude plans often already expose document skills at runtime.

Q: Scanned handwriting?
A: Expect partial OCR; require human confirmation on ambiguous fields.

Source attribution