what is speak-human-tw fr?

speak-human-tw is a skill file for AI coding assistants like Claude Code, Codex, and Cursor that edits Traditional Chinese writing to remove the telltale patterns of AI-generated text. The README, written in Traditional Chinese for a Taiwanese audience, explains that text written by AI often reads as technically correct but somehow off, full of habits a human writer would rarely use, and this tool is built to catch and fix those habits rather than to write content from scratch. It is meant for editing text you already drafted, not for generating new content. It detects more than 35 categories of AI writing patterns, things like exaggerated claims of significance, rhetorical questions used as a conclusion, overused em dashes, emoji stacking, and stock closing phrases, and it also corrects mainland Chinese phrasing and punctuation into the style used in Taiwan. By default it works in two rounds: first it lists every issue it found with the original sentence, the reason, and a suggested fix, then waits for the user to say which suggestions to accept before actually rewriting the file, so nothing gets overwritten before the user has reviewed it. The editing process follows a fixed set of steps: figure out what kind of document it is, lock a protected list of things like prices and proper nouns that must never change, decide how much to condense long text, work through each category of AI pattern, double check that no facts were altered, and for long documents score the result before deciding whether it is good enough to hand back. The project is installed by cloning it into a specific folder for whichever AI tool is being used, and after that it is triggered just by asking the assistant in conversation to remove the AI feel from a piece of text. It is released under the MIT license.