translation-quality

Orchestrates multi-agent translation with terminology verification, review, and professional typesetting.

27|6|Updated Mar 12, 2026
One-click install
npx skills add https://github.com/senshinji/claude-translation-skill --skill translation-quality-senshinji
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: translation-quality
Source: https://github.com/senshinji/claude-translation-skill
Command: npx skills add https://github.com/senshinji/claude-translation-skill --skill translation-quality-senshinji

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? Professional document translation often suffers from fabricated content, inconsistent terminology, broken table structures, and unpolished output. This Skill coordinates a team of specialized AI agents to translate documents with verified terminology, independent quality review, and publication-ready formatting. ## Core Features & Use Cases - Multi-Agent Pipeline: A translator, terminology researcher(s), and reviewer work in parallel — the researcher verifies proper nouns via web search while the translator produces a structure-preserving first pass, then an independent reviewer catches fabrication, omission, and terminology errors. - Structure Preservation: Tables stay tables, lists stay lists, and multi-item cells stay inside their cells, enforced by structural manifests and automated comparison tests. - Smart Scaling: Documents over 10 pages are chunked across parallel translators; documents with over 50 terms trigger two parallel terminology researchers (fast + deep). - Professional Output: Final translations are typeset to formal standards (Times New Roman/SimSun, A4 layout) and delivered as both .docx and .pdf. - Use Case: Ask to translate a Chinese conference agenda into English — the skill verifies organization names against official websites, preserves the schedule table exactly, reviews every paragraph against the source, and delivers formatted Word and PDF files. ## Quick Start Ask Claude Code to translate your document with a phrase like "translate this document with quality check" and the pipeline runs automatically.

Frequently Asked Questions about translation-quality

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I translate a document with verified terminology in Claude Code?

Ask Claude Code to translate the document with a phrase like "translate with quality check". The skill spawns a terminology researcher that verifies organization names, person names, and technical terms via web search, then a translator and reviewer complete the pipeline.

How does this skill prevent AI translation fabrication?

The reviewer performs a forward pass (translation to source) and backward pass (source to translation), checking every paragraph for invented content, omissions, and mismatched numbers or names. An anti-fabrication checklist defines severity levels and detection patterns.

What setup is required to use this translation skill?

Enable Agent Teams by adding CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 to the env section of ~/.claude/settings.json, then symlink the skill directory into ~/.claude/skills. No other dependencies are required.

Can it handle long documents over 10 pages?

Yes. Documents over 10 pages are split at chapter boundaries into chunks translated by parallel agents, then merged and reviewed as a whole. Documents with over 50 terms trigger two parallel terminology researchers split by search complexity.

What output formats does the translation pipeline produce?

The pipeline delivers both a Word (.docx) and PDF (.pdf) file with professional typesetting: Times New Roman for English, SimSun for Chinese, A4 pages with 1-inch margins, and minimalist table styling with gray header rows.

What happens when web search is unavailable during terminology verification?

The researcher retries failed searches once, then marks the term as low confidence with a note that web search was unavailable and continues. If all searches fail, it produces a best-judgment glossary with all terms flagged as low confidence.