
pdf-proof
Visual proof from source PDFs: highlighted values and confidence scores.
About
pdf-proof is a Claude skill built for 'trust, but check' moments with AI claims about PDFs. Given a document and a value to verify, it finds the exact text coordinates with PyMuPDF, crops the surrounding page section, highlights the match, reads the text back from the region to confirm correctness, and assembles a shareable HTML proof page with confidence scores. Scanned PDFs are handled through optional Tesseract OCR, with OCR matches capped at medium confidence. Example uses: verifying contract terms ('show me where it says the termination notice is 30 days'), tax returns, invoices, and closing disclosures. MIT licensed with 77 GitHub stars.
What it can do
- Text-coordinate search in PDFs via PyMuPDF; highlighted screenshot crops; text-readback verification with confidence scores; standalone HTML proof pages; Tesseract OCR for scanned pages; evals and example gallery.
Setup
cp -r pdf-proof ~/.claude/skills/ (Claude Code), or upload the folder as a zip under Settings > Customize > Skills (Claude Desktop)
Prerequisites
Claude Code or Claude Desktop with code execution enabled; Python with PyMuPDF and Pillow (installed automatically); Tesseract for OCR of scanned PDFs.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: MIT licensed, 77 stars, 12 commits; single skill with scripts, evals, and worked examples. Install command and OCR behavior verbatim from README. Verification mechanics (readback, confidence caps) are explicitly documented..
★ 77 GitHub stars
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.