| 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. |
Agentic Translation
Overview
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.
When to Use
Use agentic translation when ANY of these apply:
- Document size ≥ 5,000 words (multi-chunk coordination needed)
- User explicitly requests "agentic translation" (regardless of size)
- Literary/brand content with distinctive voice (style preservation critical)
- Specialized terminology requiring consistency (technical, medical, legal)
- Cultural adaptation important (idioms, references, social context)
- Maximum quality required (publication, executive, legal documents)
- Poetry/songs requiring form preservation
- Mixed formats (code + text, structured data)
- Multiple language pairs with consistency needs
Direct translation OK when:
- Simple sentence/paragraph/short text (<5K words) WITHOUT explicit "agentic" request
- User says "translate this" without mentioning "agentic"
- Style/terminology not critical
- No cultural adaptation needed
- Quick reference sufficient
Limited Context Strategy
When context is tight (< 50K tokens available):
Option 1 - Direct translation (for simple documents):
- Translate directly using your knowledge
- Apply style guidance if provided
- Review and refine
Option 2 - Subagent delegation (for complex documents):
- Create task specification (source, target, requirements)
- Dispatch subagent with the full reference:
agentic-translation-reference.md
- Subagent executes full workflow
- Collect and verify output
Quick Reference: Agent Roles
| 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 |
Core Workflow (8 Phases)
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
Critical Constraints
NON-NEGOTIABLE:
- 3 translators per chunk (never reduce)
- Style analysis required (never skip)
- Review phase required (never accept raw translator output)
- Format preservation (plain text→plain text, markdown→markdown, etc.)
- Calculated chunking (never eyeball it, includes reference sizes)
- Task tracker updated after every phase (no placeholders)
- No meta-commentary in outputs (headers/footers/synthesis reports prohibited)
- Orchestrator launches all subagents (subagents cannot launch other subagents)
- 3 translators launched in parallel (single message with 3 Task calls)
- Reviewer launched after all 3 translators complete
- Sequential chunk processing (each chunk completes before next starts)
- Condensed references for translators (full references to reviewer only)
- Cleanup required (never leave artifacts scattered)
Common Mistakes
| 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 |
Red Flags - STOP
- Fewer than 3 translators per chunk
- Reviewer launching translator subagents
- Orchestrator launching translators sequentially
- Orchestrator launching reviewer before translators complete
- Format changed (plain text→markdown, etc.)
- Meta-commentary in output files
- Task tracker contains placeholders after work begins
- Orchestrator holding document content
- Processing multiple chunks simultaneously
- Chunk files left in workspace after completion
All of these mean: Stop. Replan. Follow the workflow.
Format-Specific Guidelines
Plain text (.txt): ALL CAPS headings, preserve spacing
Markdown (.md): Preserve # headers, bold, italics
HTML: Translate text, keep tags
Code: Translate comments, keep syntax
Bottom Line
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.
