KKKKhazix/human-writing: a writing skill that targets the posture behind AI-sounding Chinese
让 AI 写的中文读起来像一个具体的人在说话。通用创作与改稿 Skill，开箱即用。
At a glance
- What is it?
- A MIT-licensed Agent Skill for Chinese prose that moves the anti-AI-flavour rules from banned strings to banned moves. It is a rule pack plus a checker script, not a model, and it only works if you actually have material to write from.
- Who is it for?
- Adopt it if you write Chinese non-fiction or fiction with an agent and you already have facts, interviews or experience to work from; the material gate is the point, not a formality. Skip it if you need English output, or if you expect a checker to fix a draft that has nothing to say.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 36 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 16, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The problem: fluent Chinese that nobody wrote
The README opens with the diagnosis in one line: AI Chinese reads smoothly, but you cannot say who wrote it. That is a different complaint from grammar or accuracy. It is about the absence of a person behind the sentences. The skill is aimed at that gap, and it is explicit about the scope: Zhihu answers, WeChat public account articles, blogs, forum posts, character stories, popular science, reviews, fiction and spoken-word scripts. In other words, most Chinese writing situations where a human voice is the product.
The intended user is someone who drives an agent to produce Chinese text and then has to publish it under their own name. The skill is not a rewriting service and not a detector. It is a set of rules the agent loads before writing, plus a script that inspects the finished draft. The repository is Python, MIT licensed, and the README states the repository contains only original rules and tools, with no third-party articles, training corpora or model weights.
The material gate comes before any style rule
Most style guides start at the sentence. This one starts earlier, with a question the README puts plainly: do you actually have something to write about. For non-fiction, the rule is to go look things up when material is thin, and if the search fails, to ask the user or shorten the piece. Padding with circular phrasing is ruled out. For fiction, invention is allowed, but every scene still needs a goal, an action and a change.
That gate is the most interesting design decision in the repository, and also the one most likely to be ignored. A skill file cannot verify that your facts are true. It can only refuse to proceed on thin material, and how strictly that refusal is honoured depends on the model reading SKILL.md. If you hand it a vague prompt and expect a 3000-word article, the honest outcome is a shorter article or a question back at you.
Three axes: material, forward motion, Chinese
Once material passes, the README describes three things being managed. The first is material discipline: for non-fiction, facts, numbers, quotations and personal experience are checked; for fiction, characters, actions and causality. The second is forward motion, stated as a rule that every paragraph must bring something new, whether a fact, an action, an example or a consequence, and must not repeat what was already written. The third is the Chinese itself: plain speech as the base, attention to word order and pauses, and removal of report tone, model tone and the reversal sentence.
The revision pass is a separate stage. The skill walks the draft paragraph by paragraph looking for places where it circles in place, cuts repeated explanations, adjusts the rhythm between long and short sentences, and blocks colon overuse, dashes, the "not X but Y" reversal construction and common AI jargon. The README is careful about the division of labour here: the checker script only enforces the hard rules that are already written down, and does not decide style for you.
1.1.0 moved the defence from strings to moves
The changelog summary in the README is the most substantive part of the whole document. Version 1.0 banned AI flavour by string: no "not X but Y", no colons, a list of banned phrases. The README says this worked, but models simply switched vocabulary and kept doing the same thing. As it puts it, "you think... actually..." and "only looking back did I realise" are the same posture as "not A but B", and readers recognise the posture, not the words.
So 1.1 bans the move itself: first planting a misunderstanding the reader never had, then knocking it down, regardless of what clothing it wears. The detection script was upgraded in parallel, adding a warning layer for disguised reversal sentences, AI parallelism and lyrical metaphor, plus statistical checks on sentence-length coefficient of variation and conjunction density. It also removed normal Chinese expressions such as "不丢人" and "打法" from the false-positive list. A distilled version under 2000 characters, dist/human-writing-lite.md, is meant to be pasted directly into chat windows such as ChatGPT or Qwen.
That upgrade is a real improvement and also a real limitation. Statistical checks on sentence-length variance and conjunction density are proxies. Prose that is genuinely uniform in rhythm, or a piece built on deliberate repetition, will trigger warnings. The README itself frames the script as a warning layer rather than a verdict.
Installing it and writing a first piece
The README's install path is a single instruction sent to your agent, which then reads the repository, locates the human-writing folder and installs it. After installation the skill displays as 「活人感写作」.
帮我安装这个skill：https://github.com/KKKKhazix/human-writingIf your agent cannot install directly, the README gives a manual route: download from Releases, or copy the human-writing folder from the repository into your local skills directory, keeping the folder name human-writing. The documented path is:
~/.agents/skills/human-writing/Once installed, the README's usage example is a prompt that names the skill and states the goal. Note that the skill is invoked by name, not by a CLI flag.
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。What you should see is a draft that first checks whether your material is sufficient, then writes, then runs a revision pass. The checker lives at scripts/check_prose.py in the repository. The README does not document its command-line arguments, its exit codes or its output format, so the only confirmed way to run it is from the repository layout itself; check the file before wiring it into anything automated. If you want the chat-window route instead of an agent install, paste dist/human-writing-lite.md, which the README describes as under 2000 characters.
Where it is the wrong tool
The skill is Chinese-specific by construction. The rules about word order, pauses, colon overuse and the reversal construction are written for Chinese prose, and forum-prose.md is scoped to Zhihu, WeChat and long forum posts. Pointing it at English output gets you a rule set that mostly does not apply.
The second boundary is harder to accept. This is a rule pack, not a model and not a fine-tune. It cannot make a draft factual, and it cannot supply a voice you do not have. If your source material is a press release and two bullet points, the skill will either shorten the piece or ask you questions, and no amount of revision will produce the texture of someone who was there. The README's own framing supports this: the material gate exists because style rules applied to empty material produce exactly the fluent-nobody problem the project is trying to fix.
Third, the checker is a warning layer. The README states it only enforces written hard rules and does not decide style. Treat its output as a list of places to look, not a pass or fail.
How it differs from a rewriting tool
The obvious comparison is a paraphrasing tool such as Quillbot, which takes finished text and swaps wording while preserving meaning. The approach here is the opposite direction of travel. Nothing is rewritten after the fact; the constraints are loaded before generation, and the revision pass works on structure, rhythm and posture rather than synonyms. A paraphraser cannot remove a reversal move, because the move survives any rewording. That is precisely the failure the 1.0 to 1.1 change was made to address.
The other comparison is a generic style guide document. A style guide is read by a human and applied by judgement. This repository is packaged for an agent: SKILL.md as the entry point, references/ split by genre and format, a dist/ file for chat windows, and a script for the hard rules. The trade-off is that the whole thing depends on the model loading and following the skill. There is no runtime that enforces it.
Maintenance, licence and what 1.1.0 costs you
The last push to the repository was on 2026-08-11, and v1.1.0 was released on 2026-08-05. The repository is not archived. There is no published roadmap beyond the CHANGELOG, and the README does not document a deprecation policy or a version compatibility matrix for the skill format, so upgrading means reading the changelog and re-reading SKILL.md.
The upgrade cost from 1.0 to 1.1.0 is not zero. The ban moved from strings to moves, and the checker gained statistical warnings, which means drafts that passed 1.0 may now raise warnings. If you have prompt templates that were tuned against 1.0's string bans, expect to revisit them. The licence is MIT, which permits commercial use and modification; the README states the repository contains only original rules and tools, with no third-party articles, corpora or model weights, which removes the most common licensing ambiguity for a writing aid. That is a description of what the repository contains, not legal advice, and if you redistribute a modified version you should read LICENSE yourself.
Editorial conclusion
Adopt it if you write Chinese non-fiction or fiction with an agent and you already have facts, interviews or experience to work from; the material gate is the point, not a formality. Skip it if you need English output, or if you expect a checker to fix a draft that has nothing to say. Before relying on it, read SKILL.md to confirm the delivery bans match your publication, run scripts/check_prose.py on one of your own finished pieces to see how many warnings it raises on prose you consider clean, and check whether your agent can install from a skill folder path rather than a one-line instruction.
Frequently asked questions
What is KKKKhazix/human-writing?
It is an MIT-licensed Agent Skill for Chinese writing that makes model-generated text read like a specific person wrote it. It ships SKILL.md as the entry point, genre references for non-fiction, fiction and forum prose, a distilled chat version under dist/, and a checker script.
What differentiates human writing from AI writing, according to KKKKhazix/human-writing?
The README frames it as posture rather than vocabulary: planting a misunderstanding the reader never had and then knocking it down is the same move whether it is phrased as "not A but B" or "you think... actually...". Version 1.1 bans that move instead of the specific strings.
How do I install KKKKhazix/human-writing?
The README says to send your agent the line 帮我安装这个skill：https://github.com/KKKKhazix/human-writing and it will read the repository and install the human-writing folder. If your agent cannot install directly, copy the folder into ~/.agents/skills/human-writing/, keeping the folder name.
Can I use KKKKhazix/human-writing in a ChatGPT or Qwen chat window?
Yes. The README says a distilled version under 2000 characters at dist/human-writing-lite.md can be pasted directly into chat windows such as ChatGPT or Qwen.
Does KKKKhazix/human-writing work for English writing?
The rules are written for Chinese prose, covering word order, pauses, colon overuse and the reversal construction, and the forum reference is scoped to Zhihu, WeChat and long forum posts. The README does not describe an English mode.
What does the checker script in KKKKhazix/human-writing actually check?
According to the README, scripts/check_prose.py enforces the written hard bans and, since 1.1, adds warnings for disguised reversal sentences, AI parallelism and lyrical metaphor, plus statistical checks on sentence-length variation and conjunction density. The README states it does not decide style for you.
Community notes