A Claude Code skill — AI translation that doesn't read like a translation.
Supports any language pair. Currently tested on English → Chinese. Feedback on other pairs welcome.
Running /translate without a target file opens maintenance mode with four options:
- Update glossary — Manage domain-specific term mappings; the more you add, the more accurate translations become
- Update translation samples — Record translation styles you like so the AI learns your preferences
- Translation discussion — Review recent translations, discuss tricky terms and style trade-offs
- Configuration — View and modify translation settings
All changes take effect immediately — no restart needed.
| Source | Typical AI Translation | Nitro Engine |
|---|---|---|
| The fact that the system is widely adopted by developers suggests that it addresses a genuine need. | 系统被开发者广泛采用的事实表明它解决了真正的需求。 | 既然这么多开发者都在用，说明它确实解决了实际问题。 |
| It is important to note that this approach may not be suitable for all use cases. | 重要的是要注意，这种方法可能不适用于所有用例。 | 值得一提的是，这套方法并非放之四海而皆准。 |
| The analysis of data was conducted using Python. | 数据的分析是使用 Python 进行的。 | 我们用 Python 分析了数据。 |
Natural & Fluent — Translation as writing, not word-for-word conversion. The output should read as if originally written in the target language.
Domain Adaptable — Custom glossaries and translation samples tailored to your industry, team, or personal preferences. The more you use it, the better it gets.
Built for Scale — Multi-agent parallel translation handles massive documents and batch folder translation with consistent style throughout.
An ablation study (GLM-5.3-class model, academic prose, with vs. without this skill's documents loaded, blind-reviewed + character-level fingerprint checks) found that current frontier models have largely internalized the translation craft this skill teaches. Given nothing but "translate this file", the model already preserved every formatting fingerprint (list markers, HTML tables, image refs, line structure, no hard-wrapping), hit the same community-standard terminology, and blind review could not reliably separate the outputs.
What still earns its place on a capable model:
- Your assets, not its rules. The user glossary and translation samples carry information — your preferred renderings, your domain's terms — which no model can internalize because it is preference, not capability. In the study, skill-loaded runs matched the user's glossary conventions 2/2; unguided runs 0/2.
- Pipeline for scale. Splitting, parallel dispatch, merge validation, and failure recovery still pay off on book-length documents and batch folders.
What no longer earns its place: the rulebook itself. Formatting discipline, code-block preservation, register awareness — and notably locale/punctuation compliance: in the study, reading the locale guide made no reliable difference to quotation-mark style, so don't count on the skill to enforce zh-CN typography. On GLM-5.3-class or newer models, treat SKILL.md as a fallback and lean on the glossary, samples, and pipeline.
Prerequisites: Claude Code + subscription plan (Claude Max or similar).
1. Install the skill
# Copy the translate/ directory to your Claude Code skills directory
cp -r translate ~/.claude/skills/2. Start translating
Translate path/to/document.md into Chinese.
Translate an entire folder:
Translate all Markdown files under path/to/folder/ into Chinese.
Translated files appear in the same directory with a language suffix (e.g., document_zh.md).
- Markdown and plain text only — PDFs and other formats need conversion first (try MinerU)
- High token usage — multi-agent parallel translation; subscription plan strongly recommended over pay-per-use
- Code blocks are preserved as-is (comments can be translated)
- Source files are never overwritten — output always has a language suffix
CC BY 4.0 — free to use, modify, and distribute, including commercially. Just keep attribution.
If you find this useful, please ⭐ the repo! This project is actively maintained and updated.