deusyu/translate-book

Rainman Translate Book – AI‑assisted whole‑book translation

What it does – Turns a PDF, DOCX, EPUB or Markdown book into a translated version in any language. The pipeline:

1. Convert the source file to Markdown (via Calibre or a layout‑aware parser).
2. Chunk the Markdown into ~6 k‑character pieces and record a SHA‑256 manifest for each chunk.
3. Translate each chunk in parallel using 8 independent LLM sub‑agents (Codex, Claude Code or OpenClaw). Each sub‑agent gets a fresh context window plus a short read‑only excerpt from the neighbouring chunks and a glossary of key terms to keep terminology consistent.
4. Validate the outputs against the manifest, then merge the translated chunks.
5. Build final artefacts – HTML with a floating table of contents, DOCX, EPUB and PDF – via Pandoc and Calibre.

Key features

• Parallel sub‑agents (default 8) avoid context overflow and speed up translation.
• Chunk‑level resumable runs; a run_state.json lets you restart without re‑translating already‑done pieces.
• Glossary system that extracts proper nouns/terms from a sample of chunks, lets you edit a glossary.json, and injects a term‑table into each translation prompt.
• Manifest‑based integrity checks (SHA‑256) guarantee that stale or corrupted chunks are not merged.
• Supports many target languages (Chinese, Japanese, Korean, French, German, Spanish, etc.) and multiple output formats.
• Optional pre‑processing for formula‑heavy PDFs using MinerU or Marker, preserving LaTeX math.

How to use

1. Install the skill for your agent runtime (npx skills add deusyu/translate-book -a codex or the equivalent for Claude Code/OpenClaw).
2. In the agent’s CLI or chat, issue a command like translate /path/to/book.pdf to Chinese (or use the slash command /translate‑book …).
3. The skill creates a temporary working directory ({book}_temp/) and, after the run finishes, you’ll find:
   • output.md – merged translated markdown
   • book.html, book.docx, book.epub, book.pdf – the final formats.

Prerequisites – Calibre (ebook-convert), Pandoc, Python 3 with pypandoc (and optionally beautifulsoup4), and an installed Codex/Claude Code/OpenClaw runtime.

Project layout – The repository contains the skill definition (SKILL.md) and a set of Python scripts handling conversion, manifest management, glossary handling, chunk context, run planning, merging and final build. Test assets are provided under tests/baselines/.

Who it’s for – Researchers, translators, or anyone who needs to translate whole books quickly while keeping terminology consistent, and who already works with LLM‑based agents.