pdf-proof

Trust, but check. A Claude skill that turns "the AI said so" into visual proof — screenshot crops, highlighted values, and confidence scores, straight from the source PDF.

When the stakes are high and the numbers have to be right — tax filings, loan documents, legal contracts, medical claims — you need more than an AI's word. pdf-proof makes Claude show its work: find the value, highlight it in the original document, verify the match, and assemble a shareable HTML proof page.

What it does

Given a PDF and a value to verify, the skill:

1. Finds the exact text coordinates using PyMuPDF's text search
2. Crops a readable section of the page around the match
3. Highlights the value with a translucent orange highlight
4. Verifies by reading text back from the highlighted region to confirm correctness
5. Assembles an HTML proof page with a summary table and per-value proof cards

The result is a standalone HTML file with embedded screenshots — a visual audit trail you can keep, share, or attach to a filing.

Examples

Tax Return — "What's the total income, taxable income, and how much is the refund?"

Lease Agreement — "What's the monthly rent, security deposit, what does it say about pets, and who pays for water?"

Mortgage Closing Disclosure — "Who's the settlement agent, what are the total closing costs, and what's the total I'll pay over the life of the loan?"

Installation

You need the entire folder (not just the SKILL.md). The script, template, and evals are all part of the skill.

Claude Desktop / Cowork

1. Download the zip
2. Open Settings > Customize > Skills
3. Click Upload skill and select the zip

The app extracts and installs it automatically. Code execution must be enabled (Settings > Capabilities).

Claude Code (CLI)

Point Claude Code at this repo and ask it to install:

"Install the pdf-proof skill from https://github.com/metedata/pdf-proof"

Or copy the folder manually:

# Global (available in all projects)
cp -r pdf-proof ~/.claude/skills/

# Project-specific (available in this repo only)
cp -r pdf-proof .claude/skills/

Skills are picked up automatically — no restart needed.

Dependencies

Python dependencies (PyMuPDF, Pillow) are installed automatically — no manual setup needed.

For OCR support with scanned PDFs, you'll need Tesseract:

# macOS
brew install tesseract

# Ubuntu/Debian
apt-get install tesseract-ocr

Use it

Ask Claude to verify values against a PDF:

"Double-check these values against my tax return and show me proof: Form 1040 line 15, Schedule D line 7"

"Verify the invoice total matches the PO"

"Show me where in the contract it says the termination notice is 30 days"

"Find the total on this scanned receipt"

The skill triggers on phrases like "confirm", "verify", "prove", "double-check", "show me proof", "where does this come from", "screenshot proof", or any request to trace a value back to a source PDF.

How it handles tricky cases

Multiple matches on the same page: A value like "3,000" might appear in both instruction text and the actual form field. The script defaults to --prefer right (rightmost match), which works well for structured forms where values are in right-hand columns. For narrative documents, use --prefer first or --match-index N.

Formatting variations: Automatically tries common variants — with/without commas, dollar signs, parentheses, periods. If a variation matches instead of the exact text, confidence drops to medium.

Scanned PDFs: With --ocr, the script detects pages without embedded text and runs Tesseract OCR automatically. OCR matches are capped at medium confidence.

Text readback verification: In verify mode, after finding and highlighting, the script reads the actual text from the highlighted PDF region and compares it against the search term. If they don't match, confidence drops to low and a warning is printed.

File structure

pdf-proof/
  SKILL.md                    # Skill instructions for Claude
  scripts/
    extract_proof.py          # Core extraction script
  assets/
    proof_template.html       # HTML template for proof pages
  evals/
    evals.json                # Test cases (replace file paths with your own)

Contributing examples

The ## Examples section at the bottom of SKILL.md is designed to grow. To add a new use case:

1. Add a new ### Category Name subsection
2. Include a "User says" prompt and "What to do" steps
3. Add any relevant "Tips" for that document type
4. Optionally add a matching eval case in evals/evals.json

License

MIT