translate

# Translate Skill Translate documents across languages using one of three modes: quick (single-pass), normal (analyze-then-translate), or refined (full four-step with polish). Core principle: **rewrite as a skilled native writer**, not word-for-word conversion. ## Reference Loading Table | Signal | Load These Files | Why | |---|---|---| | Any translation task | `references/modes.md` | Mode detection, chunking algorithm, parallel dispatch pattern | | "technical", "specialized", "glossary", "terms", or domain vocabulary in request | `references/glossary-template.md` | Glossary build, chunk injection, term-preservation rules | --- ## Phase 1: DETECT MODE AND PREPARE **Goal**: Identify mode, language pair, and document scale before any translation work. **Step 1: Infer mode from request language** | Request contains | Mode | |---|---| | "quick", "fast", "draft", "rough" | quick | | "professional", "publication-quality", "polished", "refined" | refined | | anything else | normal (default) | **Step 2: Detect language pair** - Source language: identify from content if not stated; flag ambiguity to user. - Target language: take from request; ask if absent. **Step 3: Load references** - Load `references/modes.md` for all modes. - Load `references/glossary-template.md` when the request contains "technical", "specialized", "glossary", "terms", or a domain-specific vocabulary word. **Step 4: Assess document size** - Count approximate words. - Flag documents over 2000 words ...