What is translate-academic-paper?

This project turns a PDF of an academic paper or a book, especially older scanned works, into a single self contained HTML file with a bilingual Chinese and original language view. The output page lets you switch between three views at the top: Chinese only, side by side original and translation, or the original language only. Images, formulas, and scanned pages are embedded directly in the file as base64 data, so the result has no external dependencies and can be shared as one file with anyone. The tool started as a skill for Claude Code but is designed to work with any AI agent that can read and write files and run shell commands, including other models paired with tools like Cursor, Cline, or various command line interfaces. It follows a written runbook so a different agent can pick up the process and continue it. The project sorts documents into four categories based on how clean the source is: native digital text, clean scans, messy scans with heavy OCR errors or handwritten notes, and scans with no text layer at all. For messy old scans it uses a batch process that treats OCR output as a rough guide and the scanned page images as the actual source of truth, with quality checks run after each section is translated. The tool also restores proper reading order for two column layouts, keeps technical terms and dataset names consistent using a glossary, and preserves charts and formulas as embedded images matched to the surrounding text. A firm rule in the project is that translation must be done by the agent itself, section by section, and never handed off silently to a separate machine translation service, since that has been observed to produce shortened summaries instead of full translations. Two quality check scripts, one structural and one for faithfulness, are meant to be run after every section. The project is released under the MIT license.