Open-source agent skill turns scans and phone photos into vector PDFs

On November 4, 2026, a developer open-sourced on V2EX an Agent skill for Claude Code and Codex that rebuilds scanned documents or phone photos of documents into clean PDFs where text is searchable and copyable, charts and flowcharts are vector graphics that stay sharp when zoomed in, and layout closely follows the original, rather than overlaying an OCR text layer on top of an image. To evaluate it, the author ran controlled experiments on programmatically generated test pages, each carrying a ground-truth answer set. Page types included worksheets, tables, charts, flowcharts, forms, and two-column articles, with half of the samples shot at an angle as shadowed phone photos. Every page was scored against its ground truth on text accuracy, layout fidelity, graphics quality, and vector ratio, with 1.0 as the ceiling. For each model, the same task was run twice, once without the skill and once with it, and the skill was never tuned on these pages.