pdf Editor's Pick

Our take

The pdf skill is the workhorse of the official Anthropic catalog. It does not try to be clever — it does seven concrete operations on PDF files (read, write, merge, split, OCR, rotate, fill forms) and does each of them with a confidence we wish more skills had.

This is also what makes it boring. There is nothing surprising in the SKILL.md. There is nothing novel in the approach. It is well-named, well-scoped, and well-documented. In a field where most "agent skills" are personal projects with aspirational descriptions, that boring competence is the rarest commodity.

What it does well

The trigger phrasing is textbook: "Use this skill whenever PDF files are created, read, merged, split, transformed, OCR-processed, or data-extracted." Seven verbs, one noun. An agent reading this gets a yes/no answer about applicability in one second. Compare with skills that say things like "Empower your workflow with intelligent document handling" — useless to a routing decision.

Output is also unambiguous: the skill returns a PDF file path or extracted text/table data. There is no ambiguity about what success looks like. This matters disproportionately for agent-driven workflows where ambiguous outputs cause silent failures downstream.

OCR support deserves a specific mention — it handles scanned documents through Tesseract, which means the skill works on real-world PDFs (the ones humans actually have) and not just born-digital exports. Many PDF skills assume the input is already extractable text.

What it doesn't do

The skill is honest about its limits. It does not edit PDF structure beyond what pypdf and reportlab support. Complex layout preservation when merging differently-sized pages is not handled gracefully. Form-filling supports AcroForms but not XFA (the proprietary Adobe format some legacy government forms use).

If you need precise typographic control over generated PDFs — say, a PDF that matches a specific brand guideline with exact font kerning — use a typesetting tool (LaTeX, ConTeXt, Typst) and have the agent invoke it through bash. pdf will get you 80% of the way; it will not get you to camera-ready.

When to reach for it

Reach for pdf whenever the input or output is a PDF file and the operation is one of the seven core verbs. Reach for something else when (a) you need precise layout control, (b) you are working with XFA forms, or (c) the document is born-digital and a structured format (DOCX, HTML, Markdown) would serve you better — in which case the docx or xlsx skills are usually a better fit.

Test runs (2026-05-21)

We exercised the skill against five real PDF inputs spanning typical agent use cases.

Test	Input	Result	Notes
Extract text	12-page research paper, born-digital	Pass	Plain text + page numbers preserved. Equations rendered as Unicode where possible, LaTeX-like markup otherwise.
OCR scanned page	Hand-scanned form, 300 DPI	Pass	Tesseract via pytesseract. Accuracy ~96% on typed sections, ~80% on handwriting.
Merge 4 PDFs	Quarterly reports, mixed page sizes	Partial	Output is correct but pages don't auto-fit to a uniform size. Document this for users.
Split by bookmark	A 280-page manual with 12 chapter bookmarks	Pass	Each chapter exported as a separate PDF with metadata intact.
Fill form fields	IRS W-9 PDF (AcroForm)	Pass	All standard text fields populate. Checkbox states require boolean input — works but error message could be friendlier.

What broke

Nothing catastrophic. The merge case produced an unexpected result rather than an error — the kind of thing a careful user would catch but an agent operating headlessly might not. Worth flagging in production.

Repro recipe

/plugin install pdf
# in claude code, ask:
# "extract text from /path/to/paper.pdf"
# "OCR /path/to/scan.pdf into /path/to/scan.txt"
# "merge a.pdf b.pdf c.pdf d.pdf into merged.pdf"