A Claude PDF skill is a folder of instructions that tells the agent which library to use for a PDF job and how to avoid the usual traps. The best general pick in the directory is the official PDF skill from Anthropic. Several open-licence alternatives cover the same ground, and a few specialist skills do one thing, such as converting documents to Markdown, better than the all-rounders.
This guide compares the PDF skills listed in the pdf topic hub. For each one you get a "best for" line, what it needs installed, its licence as the directory records it, the result of the automated safety check, and a trade-off. Descriptions come from each skill's own SKILL.md and repository. The editorial team did not run these skills for this comparison, so nothing below is a claim about output quality.
Which Claude PDF skill fits which job?
Most PDF requests fall into four jobs, and the skills split along the same lines.
- Everyday editing and creation: anthropics/pdf is the broadest option. hkuds/pdf, thinkinaixyz/pdf and tokenrhythm/pdf-toolkit cover similar ground with different licences and levels of structure.
- Reading a document so the agent can answer questions about it: docling-project/docling and opendatalab/mineru convert or parse files for reading, including scans.
- Producing a PDF from Markdown or HTML: shareai-lab/pdf points the agent at pandoc, wkhtmltopdf and ReportLab.
- Taking part of a file out: cherryhq/office-transform extracts selected pages into a new file and never touches the original.
If your job is a mix, pick the all-rounder first and add a reader only when you hit documents the all-rounder cannot parse well.
Best Claude PDF skills compared
| Skill | Best for | Needs | Licence | Automated check |
|---|---|---|---|---|
| anthropics/pdf (official) | General PDF toolkit | Python with pypdf, pdfplumber, reportlab | Proprietary | Passed |
| hkuds/pdf | Python sandbox work, tables to Excel | pdfplumber, pypdf, reportlab, openpyxl | Apache-2.0 | Passed |
| tokenrhythm/pdf-toolkit | Repeatable scripted operations | Python with pypdf, pdfplumber, reportlab | Apache-2.0 | Passed |
| thinkinaixyz/pdf | General toolkit with optional OCR | Python 3, same libraries, optional OCR tools | Proprietary | Passed |
| shareai-lab/pdf | Markdown or HTML to PDF, quick text extraction | pdftotext, PyMuPDF, ReportLab, pandoc or wkhtmltopdf | MIT | Passed |
| docling-project/docling | Converting documents to Markdown or JSON | Python 3.10 or newer, the docling package | MIT | Passed |
| opendatalab/mineru | Page-by-page reading with citations | The mineru command-line tool | Custom licence file, not a standard one | Warning raised |
| cherryhq/office-transform | Extracting pages into a new file | Python, run through uv | AGPL-3.0 | Passed |
"Passed" means the directory's static check reported no findings. It says nothing about how well a skill performs.
The official PDF skill from Anthropic
Best for: teams that want one skill to handle extraction, merging, splitting, rotation, watermarks, form filling and new documents.
anthropics/pdf is published in Anthropic's skills repository. The folder holds SKILL.md, a separate forms.md for form filling, a reference.md for advanced features, a set of scripts and a LICENSE.txt. pypdf covers merging, splitting and rotating, pdfplumber pulls text and tables, and reportlab creates new files. The form scripts check for fillable fields, extract field information, fill fields or annotations, and render pages to images so you can check bounding boxes.
Requirements: Python with pypdf, pdfplumber and reportlab.
Trade-off: the licence is proprietary. The LICENSE.txt states that use of the materials is governed by your agreement with Anthropic regarding its services, so read it before you copy the skill into another product or redistribute it. The directory marks the skill as official and its automated check passed.
Open-licence alternatives for general PDF work
hkuds/pdf
Best for: agents that run complete Python scripts inside a sandbox.
hkuds/pdf is licensed Apache-2.0. It picks a library per job: pdfplumber for text, tables and word coordinates, pypdf for merging, splitting, rotating, cropping, watermarking, encryption and fillable AcroForm fields, and reportlab for new documents. Flat forms are filled with an annotation overlay. Extracted tables can go to Excel with openpyxl, one worksheet per table.
Requirements: a Python sandbox with pdfplumber, pypdf and reportlab, plus openpyxl for the Excel export.
Trade-off: it is honest about scans. With no OCR engine and no network, the skill tells the agent to say the text cannot be recovered rather than guess. That is the safer behavior, but it means scanned files need a different skill. The automated check passed.
tokenrhythm/pdf-toolkit
Best for: jobs where you know exactly what you want done and prefer fixed scripts to improvised code.
tokenrhythm/pdf-toolkit is licensed Apache-2.0. The agent chooses a bundled script by goal: extract.py for text and tables, merge.py for whole files or page ranges, split.py to cut by range, form_fill.py for text form fields, and a reportlab snippet to build a PDF from data. Extraction has a table-strategy switch and a JSON output mode, and merge accepts a manifest with one-based page ranges per file.
Requirements: Python with pypdf, pdfplumber and reportlab.
Trade-off: scanned PDFs are out of scope because no OCR engine is included, and the skill points to a sibling OCR skill instead. The automated check passed.
thinkinaixyz/pdf
Best for: the same general jobs as the official skill, with optional OCR for scanned pages.
thinkinaixyz/pdf lists a proprietary licence. Its folder layout, as the directory describes it, has the same shape as the official skill: a main guide with recipes per library, a forms.md backed by form scripts, and a reference.md. It adds command-line alternatives (pdftotext from poppler-utils, qpdf and pdftk) and an OCR route through pytesseract and pdf2image.
Requirements: Python 3 with pypdf, pdfplumber and reportlab. OCR needs pytesseract and pdf2image, and the command-line tools are optional.
Trade-off: because the licence is proprietary, check the bundled terms before reusing the files. If you want permissive terms for the same kind of toolkit, hkuds/pdf is the closer match. The automated check passed.
shareai-lab/pdf
Best for: a small skill that gets text out of PDFs and turns Markdown or HTML into PDF.
shareai-lab/pdf is licensed MIT and is the shortest of the group. For reading it prefers pdftotext and offers PyMuPDF for page-by-page access with metadata. For creating files it suggests pandoc from Markdown, ReportLab in code, and wkhtmltopdf for HTML. Merging and splitting use PyMuPDF. Its best-practice list tells the agent to check that tools are installed, process large PDFs page by page, and fall back to OCR with pytesseract when extraction returns nothing.
Requirements: pdftotext from poppler-utils, Python with PyMuPDF and ReportLab, and pandoc or wkhtmltopdf for conversion.
Trade-off: form filling is not among the four jobs it covers, so use another skill if you need that. The automated check passed.
Skills for reading and extracting from PDFs
docling-project/docling
Best for: turning PDFs, including scans, into Markdown or structured JSON that the agent can read.
docling-project/docling comes from the Docling project, whose repository states an MIT licence. The skill drives the docling command-line tool for one-off local reads, a Python SDK for custom or batch pipelines, and a service client for a self-hosted or managed docling-serve endpoint. Besides PDF it handles Office files, HTML, images and other formats, and it routes scanned PDFs through OCR or a vision-language model.
Requirements: Python 3.10 or newer and the docling package.
Trade-off: it reads and converts documents. It is not the tool for filling a form or stamping a watermark. The automated check passed.
opendatalab/mineru
Best for: long documents the agent should work through page by page, with stable citations.
opendatalab/mineru tells the agent to prefer the mineru command-line tool over generic PDF parsers and OCR libraries for the formats it supports. It returns locators in a document, tier, page and block form so follow-up reads and citations point at the same place. Inputs include scanned PDFs, images, Office files, EPUB, HTML and CSV.
Requirements: the mineru command-line tool.
Trade-off: the directory lists its licence as a licence file that does not match a standard licence, and its automated check raised a warning. Open its skill page to see exactly what was flagged and read the licence file on GitHub before adopting it. That is a prompt to review, not a verdict on the project.
cherryhq/office-transform
Best for: pulling selected pages, a cell range or a slide out of a file into a new file.
cherryhq/office-transform never modifies the source: both scripts refuse to write to the source path or overwrite an existing file. For PDFs it extracts pages by one-based page number. It also handles spreadsheets, Word documents and slide decks.
Requirements: Python, run through uv.
Trade-off: the licence is AGPL-3.0, which carries sharing obligations if you build it into a networked service. The skill itself says that for simply reading a document to summarize it, a separate to_markdown tool is the right route. The automated check passed.
How to install a PDF skill
Installing any skill above follows the same pattern: copy the folder into the location your agent scans, or use the agent's plugin or CLI route. The exact steps differ by agent, so use the pages written for them: Claude Code and Codex. The guide to adding skills to Claude Code covers personal versus project scope and what to do when a skill does not load.
Before you install, do two checks. Confirm the Python libraries or command-line tools in the requirements column exist where the agent runs. Then open the skill's page and read the safety result and the SKILL.md itself, since these skills execute code on your files.
Choosing between them
Pick the official skill if its proprietary terms suit your use and you want the most widely adopted option. Pick hkuds/pdf or tokenrhythm/pdf-toolkit if you need Apache-2.0 terms, with the first leaning on agent-written scripts and the second on fixed ones. Pick Docling when the real task is reading, especially scans. Add a second skill rather than stretching one beyond what it describes.
The same document-skill family covers Office formats, and the directory compares those too in the guides to PowerPoint skills, Excel skills and Word skills. To see what else is popular, browse the top skills list.
Frequently asked questions
Which Claude PDF skill should I start with?
Start with the official PDF skill from Anthropic if you need a general toolkit for extracting, merging, splitting, filling and creating PDFs in Python. Its licence is proprietary, so if you need permissive terms, look at the Apache-2.0 and MIT alternatives in the comparison table.
Can a Claude PDF skill read scanned documents?
Some can, but only with an OCR engine installed. The official skill and one community variant mention OCR, one Apache-2.0 skill says plainly that it cannot recover text without OCR, and a dedicated reader such as Docling handles scans through OCR or a vision-language model. Check the requirements line for each skill before relying on it.
Do PDF skills work in Codex as well as Claude Code?
A skill is a folder with a SKILL.md file, and several agents load that format. Whether a particular PDF skill works depends on the Python libraries or command-line tools it expects being available where the agent runs. The agent pages on this site list where each agent looks for skills.
Are PDF skills safe to install?
The directory runs an automated static check on every skill and shows the result on each skill page. It is a first filter, not a security audit, so read the SKILL.md and any bundled scripts before you let an agent run them on important files.
What is the difference between a PDF skill and a PDF converter?
A PDF skill is a set of instructions, and often scripts, that tells the agent which library or tool to use for a job. A converter such as Docling is a tool the skill wraps. Skills decide how the agent works, and converters do the document parsing.