What the PDF skill actually gets done
The skill turns a list of PDF chores into something you can ask for in one sentence. These are the operations its own guide documents:
- Merge several documents into one, or split one into a file per page
- Pull the text of a page out, or extract tables as structured rows
- Rotate pages, add a watermark, strip or set a password
- Extract the images embedded in a file
- Fill in an AcroForm from a set of values, then flatten it
- OCR a scanned page so its text becomes selectable and searchable
Nothing here is a hosted service. The skill drives Python libraries from the terminal — pypdf for page-level surgery and form fields, pdfplumber for text and table extraction, reportlab for building new pages — and calls out to qpdf where encryption is involved and to Tesseract when a page has no text layer at all.
When it triggers — and when it doesn't
Its description is deliberately broad: use it whenever the request involves a PDF file. That covers reading a contract, pulling a table out of an annual report, splitting a scanned batch, or filling the same form twenty times.
It is the wrong tool for a few things people ask for anyway:
- PDF to Word. There is no conversion path in this skill. It can extract text and images, but it will not hand you back an editable .docx that keeps the layout. Use the docx skill for the Word side and rebuild.
- Precise layout editing. Fixed-layout pages are the point: you can drop a watermark or stamp text at coordinates, but you cannot reflow paragraphs.
- Digitally signing. It handles passwords, not certificates — there is no signing identity here, and no way to validate an existing signature.
- Forms that are not AcroForm. If the "form" is flat text on an image, filling it means drawing text at coordinates, not setting field values.
How to check it's working
Give it a two-page PDF and ask for the text of page two. If it loaded, you get the text plus the fact that it used pdfplumber or pypdf. If it did not, the agent will either apologise or invent a Python script that fails on import.
The second check is the one that saves time later: ask "is this file a scan or does it have a text layer". A loaded skill inspects the file before answering. An unloaded one assumes.
Common failures
Every page comes back empty. The file is a scan and has no text layer. Text extraction has nothing to read. Ask for OCR instead — that path needs Tesseract and Poppler installed on the machine, and the error when they are missing mentions one of those two names.
OCR runs but the output is garbage. Low resolution is the usual cause; 150 DPI is the floor and 300 is what a document scanner would give you. Skewed pages and mixed-language pages also degrade it, and there is no setting that fixes a bad source image.
Password-protected files fail to open. Pass the password in the request. Without it the library raises an error on open rather than on read, so the failure message points at the wrong place.
A form fills but nothing shows in a viewer. Page content and form field values are two separate layers. Writing values without flattening leaves some viewers showing the empty original, so ask for the form to be flattened when the output leaves your machine.
Merging loses the bookmarks or the page size. pypdf keeps pages as they are, so a merge of A4 and Letter produces a mixed-size file. If the result has to print, say so and ask for normalisation.
The agent writes its own script instead of using the skill. Off-the-shelf pypdf snippets miss the two things that actually break: encrypted inputs and rotated pages. If you see hand-written pypdf code, the skill probably did not load.
How it compares with the other Office skills
PDF is the format that arrives finished — you receive it and cannot edit it. That is why this skill is mostly read and repair work.
- xlsx is for the spreadsheet inside the report: formulas, pivot tables, and a sheet you did not build yourself.
- docx is for Word files, including tracked changes and comments, which is where a review round should happen.
- pptx is for decks built from an outline rather than edited page by page.
None of the four converts into another's format. If the job is "turn this PDF into an editable document", the honest answer is that no skill in this set does it well — extraction plus rebuild is the reliable path.
Try a sample task
This walkthrough gives you a prompt and an acceptance check to run in your own client.
Extract a table from a report
Choose a two-page PDF you can share, with a table on page 2. Keep a scanned copy for a separate OCR check.
Use the PDF skill to extract the table on page 2 to CSV. Keep the headers, check the total against the source and report any ambiguous cells.
What to verify A CSV with the expected columns and total. For scanned input, compare sampled OCR cells against the page image.
Best for Text and table extraction, scanned files, page operations and form workflows.
How to use pdf
Start with one page and check whether the PDF already has a text layer. Use OCR for image-only pages and compare the result against the original scan.
Acceptance check Compare three rows against the source, including a wrapped label and a decimal amount.
Install for your agent
Run the command from your project directory. After installation, reload the client if it has already started. Tool access and execution features depend on the client.
Claude Code
npx skills add anthropics/skills --skill pdf -a claude-codeCodex
npx skills add anthropics/skills --skill pdf -a codexCursor
npx skills add anthropics/skills --skill pdf -a cursorGitHub Copilot
npx skills add anthropics/skills --skill pdf -a github-copilotAsk your agent to install the skill
Install pdf from anthropics/skills in this project. Show me the destination folder and report any dependencies the source lists.
For manual installation, copy the entire skill folder, including references and scripts, into a skill directory supported by your client. See the installation guide.
What the upstream guide covers
The upstream PDF guide organizes its instructions by operation. It points readers toward separate approaches for text extraction, page rearrangement, form handling and OCR.
HavenSkills workflow check
November 17, 2026 · Programmatic check with pypdf on a two-page synthetic PDF.
- What we checked
- Reordered a Letter page and an A4 page, reopened the result, and verified that both pages remained and their widths were 595 and 612 points in that order.
- Scope
- This check covers page reordering. OCR, forms and an end-to-end agent run were outside the test.
Requirements, permissions and limits
- Runtime
- Python PDF libraries; OCR and page-rendering tools as needed
- File and tool access
- The agent may read every supplied page and write temporary renders or extracted text. Use a redacted sample when checking a client’s data-handling behavior.
- Limits to check
- OCR quality depends on the original scan. Multi-page tables, rotated labels and form appearance require a visual check.
Frequently asked questions
Is the PDF skill free?
It is open source under Apache 2.0, and the full terms are in LICENSE.txt inside the skill folder.
Does it upload my PDFs anywhere?
The published skill is a set of instructions and helpers. Where the PDF is processed depends on the agent and runtime you use. Check your client’s data-handling settings before using sensitive PDFs.
Can it read a scanned PDF?
If the page has a text layer, yes. If it is a scan, the text is an image until OCR runs, and OCR needs Tesseract and Poppler installed locally.
Can it turn a PDF into an editable Word file?
No. There is no conversion path. It extracts text and images; an editable Word file with the original layout means rebuilding it with the docx skill.
Can it edit pages, not just read them?
Yes. Merging, splitting, rotating, watermarking and password changes all work page by page through pypdf.