scan-to-vector 1.0

The first stable release: one command drafts every page, the agent proofreads from one sheet per page, and a benchmark shows how much it helps.

Highlights

• Every model we tested does better with the skill. On two 9-page scanned documents it had never seen, the scores went from:

  Model	Without the skill	With the skill
  Haiku 5.5	0.841	0.997
  Fable 5.1	0.837	0.997
  Sonnet 5.5	0.994	0.998
  Opus 5.5	0.993	0.998

  Sonnet also finished 37% faster. Details and charts are in the README.

• One-command draft (draft.py):

  • flattens photos and straightens pages;
  • reads the text twice and places every word at its measured spot, size and weight, in reading order;
  • traces lines, arrows and chart series into vectors;
  • compares each page with the scan.

• Proofreading sheet: doubtful words are cropped and enlarged onto one image per page. The agent fixes them by id in fixes.json.

• Install as a Claude Code plugin, or with a one-line installer for Claude Code and Codex.

• Benchmark and CI: a synthetic benchmark with ground truth (bench/), plus CI that runs the tests with Tesseract and checks the installer.

Install

/plugin marketplace add 3351666087/scan-to-vector-skill
/plugin install scan-to-vector@scan-to-vector-skill

or

curl -fsSL https://raw.githubusercontent.com/3351666087/scan-to-vector-skill/main/install.sh | bash

The scripts need Python 3.10+ with PyMuPDF, Pillow, numpy and scipy, plus Tesseract.

The attached scan-to-vector-demo.mp4 shows a phone photo of a chart rebuilt as a vector page.