Claude Translation Skill — Multi-Agent Professional Translation

A Claude Code skill that orchestrates a complete professional translation pipeline using Agent Teams.

What It Does

This skill coordinates multiple specialized AI agents working in parallel:

• Translator (Sonnet) — First-pass translation with strict structure preservation
• Term Researcher (Sonnet) — Web-verified terminology glossary with confidence scoring
• Reviewer (Opus) — Independent quality review catching fabrication, omission, and errors
• Revision — Applies all fixes and glossary corrections for the final output

User: "翻译这个文档"
         │
   Team Lead orchestrates:
   translator + term-researcher(s) [parallel]
         → reviewer [waits for both]
         → revision → typesetting → .docx + .pdf

Key Features

• Anti-fabrication pipeline: Paragraph-by-paragraph verification, no invented content
• Structure preservation: Tables stay tables, lists stay lists, cell content stays in cells
• Smart scaling: >50 terms → auto-splits into 2 parallel researchers (fast + deep)
• Web search fallback: Graceful degradation when search is unavailable
• Professional typesetting: Outputs both .docx and .pdf with formal formatting

Installation

Copy the skill to your Claude Code skills directory:

# Clone
git clone https://github.com/senshinji/claude-translation-skill.git

# Symlink into Claude Code skills
ln -s "$(pwd)/claude-translation-skill" ~/.claude/skills/translation-quality

Prerequisite: Enable Agent Teams in Claude Code:

# Add to ~/.claude/settings.json under "env":
"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"

Usage

Just ask Claude Code to translate a document:

翻译这个文档
专业翻译 meeting-guide.doc
translate with quality check

The skill auto-triggers on translation tasks involving professional documents.

Project Structure

SKILL.md                              # Main skill definition (≤500 lines)
references/
├── glossary-schema.md                # Terminology glossary JSON schema
├── review-feedback-schema.md         # Review feedback format
├── scaling-guidelines.md             # Chunking + parallel scaling rules
├── typesetting-rules.md              # Professional document formatting
├── anti-fabrication-checklist.md      # Fabrication prevention methodology
├── test-translation-structure.sh     # Structure verification test
└── test-skill-integrity.sh           # Skill integrity test (32 checks)

Testing

# Run skill integrity tests (no external dependencies)
bash references/test-skill-integrity.sh

# Run structure tests (requires source + translation files)
bash references/test-translation-structure.sh source.html translation.md

Scaling

Document Size	Translators	Term Researchers	Reviewers
≤5 pages	1	1	1
6-10 pages	1	1-2 (if >50 terms)	1
11-20 pages	2 (parallel)	2 (fast+deep)	1
21-50 pages	3-5	2 (fast+deep)	1

License

MIT