In the previous article about shuorenhua, I mentioned it’s more suitable for revising long Chinese texts.
So, is Humanizer-zh still worth discussing?
I think it is.
But its usage is a bit different.
If you directly use it to revise a whole WeChat article, it might not be the most convenient. It’s better suited for a first round of checking: exactly where does this passage feel AI-like? Is it empty background, vague attribution, forced three-point structure, or an overly positive ending?
Many people say a paragraph “has an AI flavor” and stop there.
Where exactly?
Why?
Should you delete, revise, or leave it alone for now?
Without clarifying these, later edits are just based on feeling. Editing by feeling can work, but often just changes it into another smoother AI style.
The advantage of Humanizer-zh is that it first names the problems.
See the problems first, then start editing.
1. Clarify the source first
The English humanizer repository is here:
https://github.com/blader/humanizer
The Humanizer-zh repository is here:
https://github.com/op7418/Humanizer-zh
When I looked at the source code, I confirmed one thing: Humanizer-zh clearly states its origin.
Its core files are translated from blader/humanizer, and its utility parts reference hardikpandya/stop-slop. The English humanizer is based on the Wikipedia guide “Signs of AI writing.”
So don’t treat this as a brand-new Chinese writing framework.
More accurately:
English humanizer first organized a list of English AI writing traces. Humanizer-zh translated this checklist into Chinese so Chinese users can directly use it to check texts.
It’s not complicated.
But it’s quite useful for writers. Many drafts are hard to revise not because of poor writing, but because the problems aren’t clear yet.
2. What it helps you see
For example, this passage:
Against the backdrop of rapid development of current AI tools, writing ability is becoming a key component in improving personal efficiency. By reasonably using AI flavor removal tools, creators can not only improve article quality but also better establish their own expression style.
This paragraph looks smooth.
But I feel a bit annoyed after reading it: it says many correct things but doesn’t say anything concrete.
According to the Humanizer-zh checklist, it has at least these problems:
- “Against the backdrop of rapid development” is empty background.
- “Key component” is an inflated significance.
- “By reasonably using” sounds like an instruction manual.
- “Not only… but also…” is a neat parallel structure.
If you want to revise, it doesn’t have to be fancy.
You could change it to:
I’m currently writing a WeChat public account and often let AI draft a first version.
The first draft is usually readable but too smooth. So smooth that every sentence sounds like a standard answer, but few sentences are really what I would say.
This version isn’t sophisticated.
But at least it sounds like a person talking about their own trouble.
First identify where it’s empty, then decide which sentence to delete, which subject to change, and which specific action to add.
3. How to use it
If you want to install the English version, give this address to AI:
https://github.com/blader/humanizer
Then say:
Please help me install the humanizer Skill from this repository.
If you want to install the Chinese version, give this address to AI:
https://github.com/op7418/Humanizer-zh
Then say:
Please help me install the Humanizer-zh Skill from this repository.
After installation, you can use it like this:
Please use Humanizer-zh to check the following text.
First list the AI writing patterns detected.
Then provide a more natural rewrite.
Do not add new facts or remove the core information of the original text.
[Paste your text]
If you want to quickly compare with the English original, you can ask:
Please compare humanizer and Humanizer-zh and tell me which AI writing patterns this Chinese text mainly hits.
Please explain the problems clearly, then rewrite.
For regular users, the Chinese version is enough.
For those who want to learn Skill writing, open both. You’ll see which rules can be directly migrated from English to Chinese and which need rewriting.
4. How they actually work
Let me be clear here.
I didn’t see any dedicated rewriting scripts behind these two repositories, nor did I see them calling any AI flavor removal API.
They mainly rely on the rules in SKILL.md to have AI recognize, rewrite, and review according to the checklist.
The English humanizer has more detailed rules.
It lets AI scan for these issues:
- Exaggerated meaning and trends
- Advertising-style language
- Vague attribution
- Forced three-part structure
- Common AI vocabulary
- Excessive dashes
- Overly polite and collaborative traces
- Positive but empty endings
It also has a “voice calibration” part.
If you give it a piece of your old writing, it first looks at your sentence length, word choice, paragraph openings, punctuation habits, and transitions, then rewrites according to that voice.
I quite like this.
Removing AI flavor isn’t about turning everyone into the same kind of “natural.” One person’s natural can feel awkward on another.
Humanizer-zh translates this set of rules into Chinese and adds Chinese examples and explanations.
It keeps five core principles:
- Remove filler phrases
- Break formulaic structures
- Vary rhythm
- Trust the reader
- Remove catchy lines
These five seem ordinary.
But they’re quite enough when actually revising.
For example, “remove catchy lines.” AI loves to put a sentence at the end of each section that looks like it could be screenshotted. Readers get tired of that.
Those sentences feel like they’re writing your conclusions for you, not like you really have something to say.
5. What Humanizer-zh is good for
I treat Humanizer-zh as an entry-level checklist for removing AI flavor in Chinese.
I mainly use it to break down problems.
Before, you could only say:
This paragraph has an AI flavor.
Now you can say:
This paragraph has empty background.
This paragraph has vague authority.
This paragraph is forcing three points.
This sentence sounds like an ad.
This ending is forcibly positive.
When problems have names, edits become much more specific.
For example, sentences like “research shows.”
If the original text has no source, don’t let AI make one up. A safe fix is to remove the authoritative shell and keep only the judgments you can stand behind.
Original:
Research shows that more and more creators are improving content quality through AI tools.
Can be changed to:
I do see many creators around me starting to use AI for drafting, headline editing, and data organization.
The first borrows an authority that doesn’t exist.
The second becomes your observation. The scope is smaller but more credible.
6. The difference between it and shuorenhua
If you just want to clean up a short passage, Humanizer-zh is very handy.
If you want to revise a long Chinese WeChat article, I’d prioritize shuorenhua.
The difference is clear.
Humanizer-zh is more like a problem checklist.
It tells you common issues so you can check them one by one.
shuorenhua is more like a Chinese editing workflow.
It first judges the scenario, then marks content that shouldn’t be changed, then decides the rewrite strength and sentence deletion scope. It also specially handles long Chinese texts to avoid shrinking after one edit.
So I usually use them like this:
Short texts, quick checks, learning AI writing problems: use Humanizer-zh.
WeChat long articles, technical articles, drafts needing term and fact protection: use shuorenhua.
If you want to write your own AI flavor removal Skill, open both and study.
English humanizer teaches you how to build problem categories.
Humanizer-zh teaches you how to translate and localize those categories.
shuorenhua teaches you how to put the categories into Chinese scenarios and workflows.
7. How to write a similar Skill
If you learn from these two Skills, I suggest focusing on two things: the problem checklist and the fixes.
Don’t write “make the text more natural.”
That’s too vague.
Be clear about specific issues:
Empty background
Vague attribution
Advertising tone
Forced three-part structure
Overly positive ending
No source authority
Formulaic transitions
Synonym cycling
Then provide a fix for each problem.
For example, “vague attribution”:
If there’s no clear source, don’t add one.
You can remove the authoritative shell and turn it into the author’s own observation.
If the article needs evidence, mark "source missing here" so the author can add it.
Finally, write the checking workflow.
An AI flavor removal Skill should at least have these steps:
Scan for problems
Mark detected locations
Decide if it can be changed
Protect facts and terms
Provide rewrites
Review for added facts
Review for leftover clichés
Pay attention to the last two steps.
Many rewriting Skills fail because they “don’t look back after rewriting.” They think the job is done once the sentence is smoother. The danger lies after rewriting: did it invent new facts? Did it miss key information? Did it change the author’s original judgment?
Removing AI flavor shouldn’t become plagiarism.
It’s more like editing: remove the fake, keep the real.
8. A copy-paste action to try
Try this on a draft of your own:
Please use Humanizer-zh to check the following text.
Output three parts:
1. Detected AI writing patterns.
2. Why each part feels AI-like.
3. A more natural rewrite.
Requirements:
Do not add new facts.
Do not make up sources.
Do not delete core information just to make it more conversational.
[Paste your draft]
If you want to learn Skill writing, add one more question:
Please organize the previous checking process into a reusable Skill rule.
Include trigger conditions, input, output, workflow, prohibitions, and examples.
Used this way, Humanizer-zh is more than just a revision tool.
It’s also like a negative example checklist: first see those clichés, then learn how to avoid them.
