arXiv Paper Structure-Preserving Translation Skill Released

A Claude Code skill has been released that directly reads the original LaTeX source of arXiv papers and translates them while preserving the original structure, including formulas, tables, and cross-references. To solve the problem where existing translators damage formula symbols or equation numbers during PDF conversion, it translates based on the author's original manuscript, providing results at the level of a 'translated paper'.

Key Features and Characteristics

• Structure Preservation: Section organization, numbering, formulas, tables, captions, references, and cross-references like 'Equation (3)' are maintained in the same positions as the original text.
• Multi-Agent Parallel Processing: The paper is divided into multiple chunks and translated simultaneously, sharing a glossary between chunks to ensure consistency in translating terms like 'attention'.
• Error Verification: Checks are performed by comparing the original and translated texts to catch silent errors such as missing signs (±) in tables.
• Output Formats: Supports print-ready PDF, EPUB, and DOCX, and is also available for languages other than Korean.

Installation and Usage

It is registered in the official Anthropic plugin directory and can be installed directly from claude.ai. It can also be installed using the command npx skills add kcy4334-lgtm/translate-book-arxiv -a claude-code -g. Python, Pandoc, Calibre, and Chrome (or Edge) are required, and dependencies can be checked with python scripts/doctor.py.