This skill provides guidelines for accurate, fluent translation, with agent-based parallel processing for large files and multi-file batches.
On Translation
Translation is writing across languages — rendering content expressed in one language into another, minimizing the gap between them.
Good translation goes beyond conveying information accurately. It re-expresses that information in language that is beautiful and natural, capturing the spirit of the original author. When readers encounter the translation, they should feel the piece was written in their language from the start — not carried over from a foreign one. Through parallel processing, this standard is maintained even for long texts and multi-file batches.
Core anchor: Translation is writing — not conversion, but re-expression.
Core Principles
Imagination brings something into presence; translation gives what has come into presence a voice in another language.
The Nature of Translation: Reconstruct → Negotiate → Write
Translation is not word-for-word conversion. It is a three-step process:
Step 1: Reconstruct — Build the Imaginative Space
After reading the source text, construct a complete "imaginative space" in your mind. Different types of text generate fundamentally different spaces:
- Fiction → a world of characters, scenes, and emotions that you inhabit
- Technical documentation → an expert explaining to a student, and you are listening in
- Academic paper → a rigorous argument unfolding, and you are tracking the chain of reasoning
- Blog post → a friend talking to you, sharing experience and insight
- Product copy → someone demonstrating the value of something to you
This is not simply "understanding the text" — it is immersion. It means letting the intentional content of the source text manifest in consciousness (Husserl's eidetic intuition), letting absent scenes become present in the mind (Sartre's theory of imagination), and entering the text to grasp its flowing totality (Bergson's intuition). The quality of this imaginative space sets the ceiling for the translation.
The output of reconstruction does not stay at the level of "feeling" — it must be crystallized into the "translation tone" field of the translation brief: surface features such as person and register, and deeper features such as emotional register, narrative rhythm, lexical preferences, signature expressions, the movement of thought, and a sketch of the original "voice." These fields serve as the style baseline for all subsequent translation and sub-agent alignment.
Step 2: Negotiate — Enter into Dialogue with the Source
Once inside the imaginative space, you are not a passive observer but an interlocutor who can "question the author":
- What is this word actually trying to express in this context?
- What is the closest way for my readers to understand this concept?
- Is the author's tone here serious or ironic?
- Does this metaphor have an equivalent in the target language's world?
This step is the "negotiation" phase of translation — where the translator exercises judgment and brings their own interpretive subjectivity to bear on the source. The glossary fixes unified renderings for key terms; the translation brief captures critical decisions. The more thorough the negotiation, the more fluent the writing that follows.
Step 3: Write — Compose in the Target Language (executed by sub-agents)
Writing is the ceiling of translation — when comprehension is sound and negotiation is thorough, all remaining quality differences come down to craft. The primary agent does not execute writing directly in this step; its role is to ensure the translation brief provides sub-agents with a sufficient style baseline. The five dimensions of writing (precision, rhythm, register, naturalness, locale) and their specific requirements are detailed in the "Five Elements of Writing" section of references/subagent-prompt.md.
Strategy Must Be Adaptive
Hard-coded rules cannot cover every scenario. The same "rule" may be exactly inverted across different text types:
- Passive voice: eliminate in technical docs → standard practice in academic papers
- Rhetorical devices: avoid in technical docs → essential in literary works
- Domestication vs. foreignization: user manuals favor domestication → cultural texts may call for strategic foreignization
- Cognitive load: minimize aggressively in technical docs → literary works may legitimately challenge the reader
- Linguistic economy: pursue maximum concision in technical docs → prose may require elaboration and texture
- Regional vocabulary: mainland China uses "软件、程序、网络" → Taiwan uses "軟體、程式、網路" → Hong Kong is more heavily influenced by Cantonese
Accordingly, writing craft (precision, rhythm, register, naturalness), translation techniques (domestication/foreignization, cognitive load management, long-sentence restructuring), and locale positioning (the target region's lexical system and expressive conventions) are not treated as fixed rules. They are strategic dimensions that are adaptively determined in Phase 1 based on the characteristics of the source text and the target audience. Phase 1 produces two temporary files: a glossary (_glossary.md) to standardize renderings of core terms, and a translation brief (_translation_brief.md) to capture text type, translation tone, target locale, key challenges, and targeted strategies.
Translation Workflow
Path Conventions
SKILL_DIR= the directory containing this skill (the path following "Base directory for this skill:" in the system prompt).
| Type | Location | Notes |
|---|---|---|
| Skill resources (read-only) | {SKILL_DIR}/references/ | User-persisted files that accumulate over time |
| Temporary files | Same directory as source file | _glossary.md, _translation_brief.md, _tone_sample.md |
| Split segments | _parts/ subdirectory of source file directory | Segments and manifest for large-file splits |
| Output file | Same directory as source file | {filename}_zh.md (suffix matches target language) |
Phase 0: Input Analysis
This skill handles Markdown and plain text. Other formats (web pages, PDFs, Word documents, images, etc.) must be converted to Markdown before invoking this skill.
Language detection: Source language is detected automatically. The default target language is Chinese (zh-CN); the user may specify a different target language or locale.
Scale assessment and strategy:
| Scenario | Strategy |
|---|---|
| Single file ≤ 1000 lines | Primary agent translates directly |
| Single file > 1000 lines | Split → tone anchoring → parallel → merge |
| Multiple files | Shared glossary and translation brief → tone anchoring → parallel processing |
Decision logic:
- Single file ≤ 1000 lines → proceed immediately, no user confirmation needed
- Single file > 1000 lines or multiple files → report the plan to the user (file count, total line count, translation strategy) and wait for confirmation before proceeding
Phase 1: Read-Through and Preparation
This is the foundation of translation quality. The full text must be understood before any translation begins.
- Read through the full text: Understand the topic, argument structure, and writing style to form a holistic impression.
- Load user-persisted files (paths relative to
SKILL_DIR; must be loaded using the Read tool):{SKILL_DIR}/references/user-glossary.md— the user's accumulated domain glossary{SKILL_DIR}/references/user-samples.md— the user's preferred translation style samples- Load each file if it exists; skip without error if it does not (these files accumulate gradually over time).
- Extract the glossary (write to temporary
_glossary.md):- Ground primarily in your read-through understanding, with
user-glossary.mdas a supplementary reference. - From `user-gloss
- Ground primarily in your read-through understanding, with