zh-tw-humanizer — Removing AI Tells from Traditional Chinese

AI-written Chinese has a tell. Connectives stack up, buzzwords pile on, the tone turns therapeutic, and mainland vocabulary leaks through even when the output is Traditional Chinese.

I put together zh-tw-humanizer, a portable agent skill that removes 33 AI-writing patterns in Chinese and localises the text to Taiwanese Mandarin: 影片 not 視頻, 軟體 not 軟件, 網路 not 網絡, 資料 not 數據. It is based on Wikipedia’s Signs of AI writing guide and the zh-TW AI生成文的特徵 essay.

What it covers

• 33 AI writing patterns: fake sincerity openers, counselling tone, “不是X而是Y” constructions, nominalisation (○○感/○○性), filler verbs (提升/打造/賦能), connective stacking (此外/綜上所述), em-dash abuse, quote-piling, parallelism, synonym rotation, fake authority references, chatbot traces
• 48 mainland to Taiwan vocabulary pairs: plus simplified/traditional conversion traps (干/乾/幹, 后/後, 发/發/髮), Taiwanese particle usage (喔/耶/啦/餒/蛤), and tone calibration

Available as a plain Markdown skill (works in Claude Code, pi, Cursor, OpenCode, Gemini), a Chrome extension, and a PopClip extension.