메뉴
Use when translating documents ≥5K words, or when user requests agentic translation for smaller complex documents. NOT for simple sentences/paragraphs without explicit agentic request. See agentic-translation-reference.md for complete workflow.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Use whenever the user references an image (file path, screenshot, pasted image) and you cannot see it directly. Routes to localvision's local VLM tools — no cloud, no telemetry, image bytes stay on this machine.
Use when the user provides a vague, ambiguous, or underspecified request that could go multiple directions. Also use when the user explicitly asks to enhance, sharpen, or improve a prompt before executing it. Triggers on requests like "fix the bug", "make it better", "write something about X", or any request where critical details are missing.
Use when preparing project releases with version management, changelog updates, and documentation synchronization - automates semantic version calculation, file updates, and release preparation
| name | agentic-translation |
| version | 2 |
| date | "2026-01-31T00:00:00.000Z" |
| author | Frédéric Guigand |
| description | Use when translating documents ≥5K words, or when user requests agentic translation for smaller complex documents. NOT for simple sentences/paragraphs without explicit agentic request. See agentic-translation-reference.md for complete workflow. |
Multi-agent translation system delivering world-class quality through coordinated specialization: style analysis, target language adaptation, translation memories, parallel translation (3 translators per chunk), expert review, and centralized orchestration.
WARNING: agentic-translation-reference.md contains ~4,800 words. Read only when executing multi-agent translation. For simple translations, translate directly.
Core principle: Quality at scale requires division of labor. One agent cannot maintain context, style consistency, terminology accuracy, AND cultural adaptation across large documents.
Use agentic translation when ANY of these apply:
Direct translation OK when:
When context is tight (< 50K tokens available):
Option 1 - Direct translation (for simple documents):
Option 2 - Subagent delegation (for complex documents):
agentic-translation-reference.md| Agent | Purpose | Context |
|---|---|---|
| Orchestrator | Coordinate all work, launch all subagents | NEVER runs out—central coordinator only, MUST launch translators & reviewer |
| Style Analyst | Analyze source style | Samples if large; full read if < 50K words |
| Target Stylist | Adapt style to target language | References style analysis only |
| Chunker | Split into context-safe chunks | Math: (source + instructions + condensed_refs + output + buffer) ≤ 50% context |
| Translator (×3 per chunk) | Translate chunks | One chunk each, launched in parallel by orchestrator |
| Reviewer (×1 per chunk) | Compare 3 translations, combine best | Reads 3 outputs, launched by orchestrator after translators |
Phase 0: User clarification (11 questions including format) Phase 1: Style analysis → Style analysis document Phase 2: Target style bible → Target style guide Phase 2.5: Condensed reference guides → Compact guides for translators Phase 3: Translation memories → Consistency tracker (condensed + full) Phase 4: Document chunking → Context-safe chunks (with refs in formula) Phase 5: Translation (3× per chunk) → Translation drafts Phase 6: Review (1 per chunk) → Final chunk translation Phase 7: Assembly → Complete translated document Phase 8: Cleanup and archive → Clean workspace, archived guides
NON-NEGOTIABLE:
| Mistake | Fix |
|---|---|
| Single agent for large document | Use multi-agent team |
| Reviewer launching translator subagents | Orchestrator must launch all subagents (subagents cannot launch other subagents) |
| Orchestrator launching translators sequentially | Launch 3 translators in parallel (single message with 3 Task calls) |
| Orchestrator launching reviewer before translators complete | Wait for all 3 translators to complete before launching reviewer |
| Fewer than 3 translators per chunk | 3 translators mandatory |
| Skipping style analysis | Style analysis NON-NEGOTIABLE |
| Converting format (plain text→markdown) | Output format MUST match source |
| Meta-commentary in output | Output MUST contain only translated content |
| Creating task tracker but never updating | Update tracker AFTER EVERY phase |
| Processing chunks in parallel | Sequential processing only (within chunk: translators parallel; between chunks: sequential) |
| Leaving chunk files scattered | Run Phase 8 cleanup |
All of these mean: Stop. Replan. Follow the workflow.
Plain text (.txt): ALL CAPS headings, preserve spacing Markdown (.md): Preserve # headers, bold, italics HTML: Translate text, keep tags Code: Translate comments, keep syntax
Maximum translation quality required? Read agentic-translation-reference.md and execute full 8-phase workflow.
Document ≥5K words OR user requests agentic translation? Use this skill.
Simple sentence/paragraph without "agentic" request? Translate directly. No multi-agent system needed.
Limited context? Dispatch subagent with the full reference and task specification.