Rebuild scanned pages as vector PDFs
A Claude Code and Codex skill redraws scans and phone photos as new PDFs with real text, vector diagrams and original layout, benchmarked with and without the skill. ·
Done withClaude CodeCodexClaude
- What they did
- The skill takes over the steps where agents go wrong. A script measures everything in scan pixels, one command drafts every page, and the agent reads doubtful words from an enlarged contact sheet per page. Each rebuild is overlaid on the scan, and a report lists what is still missing. Fixing a word is one line in a JSON file.
- How it went
- On two unseen 9-page synthetic documents, scores rose from about 0.84 to 0.997 for Haiku and Fable. Sonnet and Opus were already near 0.994 and reached 0.998. Gains were within noise on 3-page packets for Sonnet and Opus.
- Worth knowing
- It takes minutes of agent time per page, and Opus and Fable ran slower with the skill; for searchable text alone, OCRmyPDF is the better tool.
Try it yourself with Claude Code
I have scanned pages and phone photos in [folder path]. Rebuild each one as a new PDF with real selectable text, vector redraws of any diagrams or figures, and the original layout, fonts and spacing matched as closely as you can. Save the results in [output folder] without touching the originals, then compare each new PDF with its source and list any spots where text or diagrams differ so I can check them.
Want help getting this working? Join the Vibe Coding Academy ↗ Run by Alex Finn, who also runs this site.
Source: HN · Nov 4, 2026
Related use cases
Build a C compiler with a team of parallel agents
Anthropic's engineering write-up on tasking Opus 4.6 agent teams with building a C compiler.
Triage email by voice while walking
Allie Miller used ChatGPT voice mode with Codex to triage important emails, draft replies, archive or delete items and log action items.
Automate copy-paste and backend wiring work with Codex
A commenter says they use Codex for work that used to involve lots of copy-pasting and backend wiring.
Run a solo company with eight agent departments
A founder set up CEO, CFO, COO, lawyer, accountant, marketing, CTO, and improver agents that share memory and consult each other.
Build a 52-node n8n workflow
Claude computer use built a 52-node n8n workflow in one afternoon that would normally take three weekends.
Extract supplier invoice PDFs into a monday.com board
A Make workflow watches a Google Drive folder for supplier invoice PDFs, extracts data with AI Extractors, and writes it into monday.com.
Get 5 like this every morning
The best things people got an AI agent to do, each with the prompt to try it.