Chinese Technical Writing

Chinese Technical Writing

Core Contract

Reader-facing Chinese research documents must use Chinese as the default technical language. English is a precision tool for names, acronyms, symbols, searchability, and source fidelity; it is not a substitute for writing clear Chinese prose.

Use this skill for Chinese Feishu/Lark pages, Notion pages/databases, Obsidian/LLM Wiki notes, paper cards, daily reviews, meeting notes, paper deep dives, source translations, summaries, figure captions, tables, and TODO/review documents. Use research-doc-workflow when the prose will be written into a durable research document.

Allowed English

Keep English when it is genuinely a name, identifier, or source-fidelity object:

• paper titles, method/model/system names, dataset/benchmark names, code/repo names, organization names, and product names;
• named architectures, model families, pretrained models, and named components such as Transformer, DINO, CLIP, NeRF, LoRA, SAM, and Gaussian Splatting; keep the official English name even when a Chinese translation exists;
• standard acronyms and symbols such as SMPL-X, 3DGS, CLIP, PSNR, LPIPS, FID, FPS, GPU, API, $L_2$, and $\mathcal{L}$;
• code identifiers, file paths, commands, config keys, CLI flags, class/function names, and quoted error messages;
• exact original English text in 英文原文稿, bilingual source archives, citations, block quotes, paper abstracts copied as source text, and other source-preserving sections;
• References / bibliography entries in paper translations and deep dives: keep the original English authors, titles, venues, publisher names, page ranges, DOI/arXiv strings, and other reference fields instead of translating them into Chinese;
• required bilingual mind-map nodes or headings when another workflow explicitly requires English plus Chinese.

Chinese-First Rule

For ordinary Chinese prose, translate generic technical concepts into Chinese. Do not translate official names of architectures, models, methods, datasets, benchmarks, systems, or components. If a retained English name needs explanation, add a short Chinese gloss on first use, for example Transformer（基于自注意力的网络架构）; do not mechanically turn the name into a Chinese phrase. If an English term is generic rather than a name and helps search or disambiguation, write it only on first use as:

中文术语（English term）

After the first occurrence, use the Chinese term or a stable acronym. Do not leave raw English noun phrases inside Chinese sentences when they have clear Chinese equivalents.

Good:

• 参数化人体估计（parametric human estimation）需要同时约束人体形状、姿态和相机位姿。
• 透视畸变会让近处身体部位被放大，从而影响人体姿态估计。
• 场景几何决定了可行接触、遮挡关系和运动边界。

Avoid:

• parametric human estimation 需要处理 pose 和 camera。
• 这个方法主要解决 perspective distortion。
• 模型没有显式利用 scene geometry。
• Transformer 和 DINO 应保留为官方名称；可以写成 Transformer架构 或 DINO预训练模型，但不要替换成非官方中文名称。

Preferred Translations

Use Chinese for common translatable terms unless they are part of an official name:

If a term has no stable Chinese translation or the English form is the community standard, keep English but add a short Chinese explanation when first introduced.

Document-Specific Handling

• Paper cards: also use paper-card-delivery; the card's metadata takeaway, caption, and seven fixed bullets must be Chinese-first.
• Paper deep dives: preserve official source text in 英文原文稿; enforce Chinese-first wording in the body of 原文中译稿 and parent-page 精读稿; keep References / bibliography entries in the original English rather than translating them.
• Bilingual source archives: do not modify the English original blocks; enforce this rule only on Chinese translation blocks or Chinese notes.
• Feishu rich pages: normalize wording with narrow text edits; do not rewrite or damage images, grids, tables, formulas, or whiteboards just to fix language.
• Notion pages/databases: preserve native blocks, properties, relations, equations, captions, and page links while editing prose.
• LLM Wiki / Obsidian notes: use Chinese body text, 中文（English term） on first use, LaTeX for formulas, and English only for names/identifiers.
• Diagrams and tables: Chinese labels by default; keep method names, dataset names, symbols, and required bilingual node formats unchanged.

Audit Checklist

Before delivering a Chinese document, check:

• Chinese prose does not contain avoidable English phrase islands.
• First-use English parentheses are used only when useful, not after every term.
• Method/model/dataset/code names, acronyms, formulas, and source quotes are preserved.
• Repeated generic English words such as baseline, ablation, metric, benchmark, pipeline, and framework have been replaced with clear Chinese unless they are official names. Named architectures and models such as Transformer and DINO remain in English.
• Captions, bullets, summaries, TODOs, and interpretation sections read naturally in Chinese.

Validation Script

For local Markdown or text drafts, run:

python .tools/skills/chinese-technical-writing/scripts/validate_chinese_terms.py path/to/draft.md

The script checks Chinese-containing lines for known translatable English phrases. It is a lint helper, not a complete language judge. If no local draft exists, do the audit manually before writing to Feishu or answering the user.