human-writing README

The project's README from github.com/KKKKhazix/human-writing. As one markdown file: README.md.

活人感写作

Version 1.1.0 MIT License GitHub Release

快速安装 · 写作流程 · 仓库结构 · 提交问题

AI 写中文有个通病:读完觉得挺流畅,但说不出是谁写的。活人感写作想治的就是这件事。

让模型写出来的文章读起来像一个具体的人在说话——知道一些事,有判断,偶尔岔开一句,还能接回来。适用于知乎回答、公众号文章、博客、论坛帖、人物故事、科普、评测、小说、口播等大多数中文写作场景。

它做什么

写作之前先解决一个前置问题:你手上有没有东西可写。

现实题材,材料不够就去查,查不到就追问或者缩短篇幅,绝不拿车轱辘话凑字数。虚构题材可以自由创造人物和情节,但每个场景仍然要有目标、有动作、有变化。

材料过关之后管三件事:

材料 推进 中文
现实写作核准事实、数字、引语和亲历。虚构写作检查人物、行动与因果。 每段都要带来新东西——新事实、新动作、新例子或新后果。写过的不重复。 白话打底,在意词序和停顿,清掉报告腔、模型腔和翻案句。

初稿写完还有一道关。Skill 会逐段检查有没有在原地转圈,砍掉重复解释,调整长短句节奏,拦住冒号滥用、破折号、「不是……而是……」之类的翻案腔和常见 AI 黑话。检查脚本只管已经写明的硬规则,不替你决定风格。

快速安装

把下面这句话发给你的 Agent。

帮我安装这个skill:https://github.com/KKKKhazix/human-writing

Agent 会读取仓库、找到 human-writing,完成安装。装好之后显示名为「活人感写作」。

Agent 不支持直接安装时

从 Releases 下载,或者把仓库里的 human-writing 文件夹完整复制到本机 Skills 目录。文件夹名保留 human-writing。

~/.agents/skills/human-writing/

装好之后这样用:

使用 $human-writing,把我的材料写成一篇有活人感和中文韵律的作品。

1.1.0 改了什么

1.0 用字符串禁令拦 AI 味——禁「不是……而是……」、禁冒号、禁一批黑话。有效,但模型会换一套字面继续做同样的事。「你以为……其实……」「回头才发现」和「不是A而是B」是同一个姿势,读者认的是姿势,不是字。

1.1 把防线从字面挪到动作:禁的是「先给读者立一个他没有的误解,再推翻它」这件事本身,不管穿什么外衣。检测脚本也跟着升级,补了变形翻案句、AI 排比、抒情借喻的警告层,加了句长变异系数和连词密度的统计检查,同时把「不丢人」「打法」这类正常中文从误伤名单里捞出来。另外出了一个两千字的蒸馏版,ChatGPT、千问这类聊天窗口直接粘贴就能用。

完整变更见 CHANGELOG.md。

仓库结构

展开查看完整目录
human-writing/
├── SKILL.md
├── VERSION
├── LICENSE
├── agents/
│   └── openai.yaml
├── dist/
│   └── human-writing-lite.md
├── references/
│   ├── forum-prose.md
│   ├── reality.md
│   ├── fiction.md
│   ├── formats.md
│   └── revision.md
└── scripts/
    └── check_prose.py
位置 干什么的
SKILL.md 入口。材料门槛、现实与虚构分流、写作流程、交付禁令,全在这一份里
forum-prose.md 知乎、公众号、论坛长帖的写法,节奏和措辞的具体做法都在这里
reality.md 真人、历史、新闻、数据和个人经历的事实边界
fiction.md 小说、故事、虚构散文和对白的创作规则
formats.md 短内容、口播、演讲、教程、评测等特殊形式
revision.md 初稿写完之后怎么改——逐遍检查清单
check_prose.py 检查成稿有没有踩到硬禁令
human-writing-lite.md 蒸馏版,两千字以内,聊天窗口直接粘贴用

反馈

MIT 协议开源。仓库只有原创规则和工具,没有第三方文章、训练语料或模型权重。

碰到规则冲突、误报或者某个模型上表现不对,欢迎提 Issue。附上你的提示词、模型输出片段和你觉得应该是什么样,排查起来快很多。

活人感写作 · Human Writing · 1.1.0

human-writing (KKKKhazix/human-writing)

claude-skill
Searching Database…

What is this?human-writing is an Agent Skill that makes AI-generated Chinese text read like a specific person wrote it instead of generic AI-sounding prose. It covers most Chinese writing scenarios: Zhihu answers, WeChat articles, blogs, forum posts, profiles, explainers, reviews, fiction, and spoken scripts.
Is this a coding library, or an instruction set for AI agents?It is not code to import. It is a Skill folder (SKILL.md plus topic-specific guide files and a Python checker script, checkprose.py) that an AI agent reads and follows, or that you paste directly into a chat window as a condensed prompt.
How do I install the skill?Tell your AI agent to read this repository and install human-writing; it will find and set up the skill automatically, showing up as '活人感写作'. If your agent can't install directly, download it from Releases or copy the human-writing folder into your local Skills directory, keeping the folder name unchanged.
Which AI agents or chat tools can use this skill?Any agent that supports the Skills format can install it directly. For plain chat windows like ChatGPT or Qwen (千问), there's a condensed under-2000-character 'lite' version (human-writing-lite.md) you just paste in.
Does the checker script guarantee good writing style?No. checkprose.py only catches hard-coded rule violations like banned rhetorical reversals, AI-ish clichés, and punctuation abuse — it doesn't make broader style decisions for you.
What changed between version 1.0 and 1.1?1.0 banned specific AI-sounding phrases, but models just swapped the wording while doing the same rhetorical trick. 1.1 instead bans the underlying move (setting up a false expectation then reversing it) and adds statistical checks for sentence-length variation and conjunction density.
How is the guidance organized across the repo?Separate files handle different writing modes: forum-prose.md for long-form posts, reality.md for nonfiction fact-checking, fiction.md for fiction rules, formats.md for short/scripted content, and revision.md for the final pass-by-pass editing checklist.
How does it decide if there's enough material to write from?For nonfiction, if the material is thin it searches for more, asks follow-up questions, or shortens the piece rather than padding with filler. Fiction can invent freely, but every scene still needs a goal, an action, and a change.
Who is this aimed at?Chinese-language writers and anyone using AI agents to draft Zhihu answers, WeChat essays, blogs, forum posts, profiles, reviews, fiction, or voiced scripts who wants the output to not read as obviously AI-written.
Creator
KKKKhazix
Size
0.5 MB
Stars
4.1K
Age
2mo
Commits · 90 Days