Document workClaude apps and Claude Code

Anthropic / pdf

What can the pdf skill do that Claude cannot do alone?

File operations. Reading a PDF is one thing; rebuilding one is another.

What to remember

  1. OCR on scans is the capability people most often miss, and the one that unlocks old archives.
  2. Merging, splitting and rotating are mechanical wins that remove a manual tool from the workflow.
  3. Table extraction is the highest risk step. Verify a sample before trusting a batch.
  4. Encrypted files need the password. The skill does not remove protection you do not have rights to.

01What it changes in practice

A PDF is a layout format, not a data format, which is why extraction is the hard part. The skill turns a category of work that usually needs a separate desktop tool into something an agent can do inside the same conversation as the analysis.

The assembly direction matters just as much. Producing one clean PDF from many sources, with pages in the right order and a watermark applied, is a small task that consumes a surprising amount of human time.

  • Extract text and tables, including from scanned pages through OCR.
  • Combine, split, rotate and watermark existing files.
  • Fill forms and handle encrypted documents where you hold the password.

02Where it disappoints

Multi column academic layouts, tables that span pages and low quality scans still produce reading order mistakes. The failure is quiet, because the extracted text looks fine until you compare it with the page.

Treat any extraction that feeds a decision as a draft that needs a sampled check, especially in finance and compliance work.

Install and first run

  1. Install the document skills from anthropics/skills.
  2. Start with one real file and one narrow operation, for example extracting the line items from an invoice.
  3. Check the extraction against the original page before you build anything on top of it.

Questions people actually ask

Can it read scanned documents?
Yes, through OCR, which is what makes scanned archives searchable rather than just viewable.
Can it fill in a PDF form?
Yes, form filling is one of the supported operations, along with merging, splitting and watermarking.
Is table extraction reliable?
It is the least reliable part. Verify a sample against the source pages before trusting a large batch.

Sources

  1. anthropics/skills on GitHubrepository
  2. anthropics/skills, pdf skill directorydocumentation

Machine readable versions of this page