wtf is human-writing?

This project is a writing skill for AI agents, meant to make Chinese text written by an AI read like it came from a real, specific person instead of a generic chatbot. It targets common Chinese writing situations: Zhihu answers, WeChat public account articles, blog posts, forum threads, biographical stories, science explainers, reviews, short fiction, and podcast scripts. The idea is that before writing anything, you need real material to work with. For nonfiction topics, if you do not have enough facts, the skill is meant to go look them up rather than pad the piece with filler or vague, rambling sentences. For fictional topics, it can invent characters and plot, but every scene still needs a clear goal, real actions, and some kind of change happening. Once material is gathered, the skill checks three things: whether the facts, numbers, and quotes are accurate for nonfiction, or whether the characters and cause and effect make sense for fiction, whether every paragraph brings something new so nothing repeats itself, and whether the tone sounds like plain, natural speech rather than a canned AI report voice or an obviously translated sentence. Before a draft is considered done, the skill runs one more pass to catch things that circle back to the same point, cut repeated explanations, adjust sentence rhythm, and remove overused colons, certain punctuation patterns, and stock AI phrasing tics. The checking script only enforces rules that are already written down, it does not decide the writing style for you. Installation is done by handing your agent a short command that points at this GitHub repository, and the agent reads the repo and installs the skill automatically. If your agent cannot install directly, you can download it from the Releases page or copy the skill folder into your local skills directory by hand. Once installed, you use it by asking the agent to turn your material into a piece with natural, human sounding rhythm. The repository is released under the MIT license and contains only original rules and tooling: no third-party articles, training data, or model weights are included.