Reach for the pdf skill instead of pasting pdfplumber boilerplate into every prompt
A bundled skill that handles the five things people actually do with PDFs — extract text and tables, merge, split, fill forms, and OCR scans — and triggers the moment the word "PDF" shows up in a prompt.
The pdf skill is one of Anthropic's bundled skills — a folder of helper scripts, references, and prompts Claude loads the moment a PDF shows up in the conversation. It covers the five operations that account for almost every PDF task at the command line: extract text and tables, create new PDFs from content, merge or split existing ones, fill and read forms, and OCR scanned pages. The trigger surface is intentionally broad — the words PDF, .pdf, form, extract, merge, or split in a prompt are enough — so you almost never have to invoke it explicitly. Claude reads the SKILL.md, picks the right helper based on the shape of the file, and runs it. You get back text, a CSV, a merged file, or a JSON of form fields, depending on what you asked for.
Why this is the right Skill pick today. PDFs are the format every enterprise workflow gets stuck on. Vendor reports arrive as PDFs, AWS billing exports get re-rendered to PDF for finance, compliance attestations are PDFs, and scanned receipts from a team that still scans receipts are PDFs that aren't even text underneath. Without the skill, you end up pasting a thirty-line pdfplumber snippet into the prompt every time, or worse, asking Claude to write one fresh and hoping it guesses the column boundaries right. With the skill loaded, the boilerplate is gone — the helper scripts already handle the common shapes: multi-page tables, rotated pages, encrypted files, fillable forms with named fields.
Pull text and tables out of a vendor report. The most common entry point. Drop a PDF in the working directory and ask: “extract the cost tables from billing-may.pdf into CSV.” The skill picks the right parser based on the file's structure — born-digital PDFs go through the fast text-extraction path, scanned pages get routed to OCR automatically. The output is a real CSV with header rows preserved, not a wall of text you then have to re-parse by hand. For multi-page tables that continue across page breaks — the most failure-prone case in any PDF pipeline — the skill knows to stitch the rows back together rather than emitting one CSV per page and leaving you to glue them.
Merge or split when you have a stack to wrangle. Combining three postmortem PDFs into one packet for the incident commander, or pulling pages 14–22 out of a 200-page compliance binder, are both one-line prompts. The skill is wired to pypdf under the hood, so it preserves bookmarks, links, and form fields across operations rather than flattening the document into a re-rendered shell. The same path handles rotation, page reordering, and watermarks — enough to cover most “I need to send this to legal but page 47 is sideways” moments without you opening Acrobat.
Fill forms and read forms. The surface most people don't realise is there. The skill can read a fillable PDF — a vendor security questionnaire, an HR onboarding form, a SOC 2 evidence template — and dump every field name and current value to JSON. It can also do the inverse: take a JSON payload and emit a filled, optionally flattened PDF. Pair that with a single JSON file you keep up to date once, and you stop re-typing the same nine answers (legal entity, DPO email, data residency, encryption-at-rest posture) into every vendor security review that lands in the queue.
The trap to avoid is reaching for the skill when you have a structured source upstream. If the data already exists in HTML, in an API response, or in a CSV from the system that generated the PDF, scraping it out of the downstream PDF is strictly worse — slower, lossier, and dependent on a layout that can change without warning. The pdf skill is for cases where the PDF is the source of truth. When something else is, fix upstream and skip the skill entirely.
See the try-it block for a one-prompt invocation against whatever PDF is sitting in your downloads folder right now.
Drop any PDF in your working folder — a vendor invoice, an AWS cost export, a scanned receipt, last quarter's compliance packet — and paste this into a Claude Code session or Cowork chat that has the pdf skill available:
Use the pdf skill on invoice-may.pdf. Extract every table to a
single CSV called invoice-may.csv, then tell me which columns
look like currency amounts so I can double-check the parser
didn't drop the $ or € prefix on any of them.If the file is a scan rather than a born-digital PDF, swap the filename and add one sentence: “This is a scanned image, so OCR it first.” The skill picks the OCR path automatically based on the file's structure, but the explicit hint keeps it from trying the fast extractor first when you already know the answer. For a merge or split, the equivalents are equally short — “merge postmortem-1.pdf, postmortem-2.pdf, postmortem-3.pdf into one file packet.pdf” or “pull pages 14 through 22 of compliance-binder.pdf into excerpt.pdf.” Same skill, same trigger, no boilerplate.