Translate Book (Parallel Subagents) Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection. A Claude Code skill that translates entire books (PDF/DOCX/EPUB) into any language using parallel subagents. Each chunk gets an isolated context window — preventing truncation and context accumulation that plague single session translation. Pipeline Overview Prerequisites Verify all tools are available: Installation Option A: npx (recommended) Option B: ClawHub Option C: Git clone Usage in Claude Code Once the skill is installed, use natural language inside Claude Code: The skill orchestrates the full pipeline automatically. Supported Languages Code Language zh Chinese en English ja Japanese ko Korean fr French de German es Spanish Language codes are extensible — add new ones in the skill definition. Running Pipeline Steps Manually Step 1: Convert to Markdown Chunks This produces inside {book name} temp/ : chunk0001.md , chunk0002.md , ... (source chunks, ~6000 chars each) manifest.json (SHA 256 hashes for validation) Step 2: Translate (Parallel Subagents) The skill handles this step — it launches 8 concurrent subagents per batch, each translating one chunk independently: Resumable: Already translated chunks (valid output chunk .md files) are skipped on re run. Step 3: Merge and Build All Formats Before merging, validation checks: Every source chunk has a matching output file (1:1) Source chunk hashes match manifest.json (no stale outputs) No output files are empty Outputs produced: File Description output.md Merged translated Markdown book.html Web version with floating TOC book.docx Word document book.epub E book format book.pdf Print ready PDF Project Structure How Manifest Validation Works If validation fails, the script auto deletes stale output.md and re merges from valid chunk outputs. Real World Example: Translate a Technical Book Resuming an Interrupted Translation Changing Output Metadata After Translation If you need to update the title, author, template, or image assets without re translating: Do NOT delete chunk files — those are your translated content. Only delete final artifacts when changing metadata. Troubleshooting Problem Solution Calibre ebook convert not found Install Calibre; ensure ebook convert is in $PATH Manifest validation failed Source chunks changed — re run convert.py Missing source chunk Source file deleted — re run convert.py to regenerate Incomplete translation Re run the skill — resumes from last valid chunk Changed title/template but output unchanged Delete output.md , book .html , book.docx , book.epub , book.pdf then re run merge and build.py output.md exists but manifest invalid Script auto deletes stale output and re merges PDF generation fails Verify Calibre has PDF output support; try ebook convert help Empty output chunks Retry failed chunks; check API rate limits Diagnosing Chunk Issues Configuration Tips Chunk size: ~6000 chars per chunk is the default. Smaller chunks = more parallelism but more API calls. Concurrency: Default is 8 parallel subagents per batch. Adjust in SKILL.md if hitting rate limits. Languages: Add new language codes to the skill triggers and translation prompt in SKILL.md . Templates: Customize scripts/template.html and scripts/template ebook.html for different HTML/ebook styling. Key Design Principles 1. Isolated context per chunk — each subagent starts fresh, preventing context overflow on long books 2. Hash based integrity — SHA 256 tracking catches stale or corrupt translated chunks before merging 3. Resumable at chunk granularity — never re translate what's already done 4. Format agnostic input — Calibre handles PDF/DOCX/EPUB normalization before the pipeline begins 5. Multiple output formats — single pipeline produces HTML, DOCX, EPUB, and PDF simultaneously

translate-book-parallel