Humanizer: Make AI-generated text more like what humans wrote

As big language models become more popular, AI writing has become a part of daily work. Writing emails, writing blogs, organizing documents, making product introductions, and even writing technical descriptions, many tasks can be handed over to AI.

But an increasingly obvious question also emerged:The words written by AI are often "right", but it is easy for people to see at a glance that it was written by AI.

The sentence patterns are too regular, they prefer to use fixed three-paragraph expressions, frequently emphasize "important meaning", they use a lot of high-frequency AI words such as "delve","testament" and "landscape", and each paragraph habitually summarizes with one sentence-These details when added together will create an obvious "AI flavor".

Humanizer is an open source Agent Skill designed to solve this problem.

1. What is a Humanizer?

Humanizer is an open source project on GitHub with the repository address: blader/humanizer。

Its core goals are very direct:

Without changing the original intention, change the words that sound like AI wrote to be more like people wrote.

Unlike traditional "AI text detectors", Humanizer does not try to determine whether a paragraph of text was generated by AI, but directly participates in text modification.

More accurately, it breaks down the "AI flavor" into a series of writing patterns that can be observed and processed, and then lets the language model re-examine the article according to these patterns.

The project itself uses Skills in Markdown format, so it can theoretically be used by Agents that support Skills, and is not limited to a specific AI product.

2. Why does AI text tend to have a "machine flavor"?

An important point behind Humanizer is:

Language models tend to select the expressions that are statistically most likely to occur, while human authors usually write in specific contexts, specific audiences, and specific experiences.

This leads to two different writing tendencies.

AI often likes:

• Use expressions that are valid for most readers;
• Emphasize the importance of something;
• Use sentences that are very well-structured;
• Frequent use of similar connection methods;
• End the paragraph with a concluding sentence;
• Create a "sense of integrity" through format and rhetoric.

Real human writing is often not so neat.

People may suddenly change a sentence pattern, may only write a very short sentence, or they may leave an expression that is not "perfect".

This characteristic of imperfect rules instead forms part of natural language.

Humanizer's idea is to find out these common "regular expressions" and consciously break them.

3. 25 models that Humanizer focuses on

The current project README organizes these writing patterns into 25 categories, and organize them according to intensity and frequency of occurrence.

These patterns can be roughly divided into five groups.

1. Replace direct expression with "foreshadowing"

The first group focuses on whether the article has added unnecessary foreshadowing to create momentum.

For example:

"It's not just X, it's Y."

Or:

"What's really important is not…but…"

This structure is completely normal to be used occasionally, but if it appears repeatedly in an article, it will easily create a template feeling.

Humanizer tends to say Y directly.

Similar questions include:

• Add a "golden sentence" at the end of each chapter;
• Use profound expressions such as "essentially" and "ultimately";
• Use sentences such as "Let's dive in" as the beginning of the formal content;
• Create a "hypothetical opposition" in order to refute a point that no one actually puts forward.

The core principles can be summarized in one sentence:

If a sentence does not add facts or opinions, consider deleting it.

2. The rhythm of mechanization

The second group targets the "neat rhythm" that AI particularly likes.

For example:

"Innovation, inspiration and insight."

A row of three words is very common.

The problem is not the "three" per se, but the fact that AI often mechanically organizes information into three projects.

In addition, Humanizer also focuses on:

• Multiple consecutive sentences use the same beginning;
• Abuse of dashes;
• Excessive stacking of qualifiers;
• A large number of hyphens are used to form compound words;
• Excessive use of passive voice;
• Omit the subject that should really appear.

Looking at these expressions individually does not mean that the text must be written by AI, so the project also emphasizes that some patterns should only be modified when multiple signals appear at the same time.

This is a more important design idea:

Humanizer does not change all sentences that "don't follow the rules".

4. The third type of problem: writing ordinary facts too important

This is a very common phenomenon in AI texts.

For example, an ordinary event:

"The project will be released in 2025."

AI may be further written as:

"This release marks an important milestone in the development of the project."

But without additional factual support,"important milestones" actually add no information.

Humanizer calls this phenomenon Inflated significance"Meaning expansion."

Similar expressions include:

• “marking a pivotal moment”
• “continues to thrive”
• “the future looks bright”
• “a testament to……”
• “showcasing……”

These words are not necessarily wrong, but if the article uses them constantly, the text tends to become empty.

Humanizer handles it as follows:

Retain facts and delete meaningful judgments that are not supported by evidence.

This is especially important for technical documentation.

5. Avoid "borrowed sense of authority"

Another issue worth noting is that AI likes to use vague authoritative sources to enhance the credibility of articles.

For example:

"Experts generally believe that..."

Or:

"This view has been widely recognized by the New York Times, BBC, Financial Times and other media."

If it does not clearly indicate which expert, which article, and what specifically said, this expression does not actually provide much effective information.

Humanizer calls this model Borrowed authority。

It tends to require:

• Give the true source;
• Explain what the source specifically said;
• Or simply delete authoritative descriptions that cannot be verified.

This also reflects an important principle of Humanizer:

It should not fabricate information to make the article seem more natural.

6. The format itself may also produce "AI flavor"

Many people think that the flavor of AI comes only from vocabulary, but in fact, format is equally important.

Humanizer pays special attention to:

Excessive use of bold

For example:

Performance: Performance improved.

This format looks like a knowledge article automatically generated by AI.

Decorative title

For example:

🚀 Launch Phase: Strategic Negotiations And Partnerships

Humanizer tends to use more natural titles and reduce emoji, arrows, and overly formal title formats.

Unnecessary curved quotes

Projects also list curved quotes as a weak signal and recommend handling them based on specific text styles.

These rules together explain one thing:

The so-called "AI flavor" is not just a few keywords, but a style formed by vocabulary, sentence pattern, structure and typography.

7. The most valuable point of Humanizer: it's not just a "one-click touch"

Humanizer's workflow is not simply about letting the model execute:

"Please write this passage more like a person."

Instead, it requires the model to first identify specific problems in the text.

According to the project README, it will:

1. Mark discovered AI writing characteristics;
2. Rank these issues according to their strength;
3. Redraft the text without adhering to the original structure;
4. Check the rewritten content;
5. Compare the original text to confirm that facts and opinions have not been changed;
6. Output the final version.

This process is actually very close to a human editing workflow.

In other words, Humanizer is not simply pursuing "more human", but is trying to answer:

"Why doesn't this sentence sound like a specific person would write this?"

This is more in-depth than simply replacing a few high-frequency AI words.

8. The most important constraint: cannot be fabricated for nature's sake

This is a point worthy of special attention in Humanizer design.

The project clearly requires:

• Names cannot be created out of thin air;
• Numbers cannot be created out of thin air;
• Dates cannot be created out of thin air;
• Quotations cannot be created out of thin air;
• Sources cannot be created out of thin air;
• Facts that are not present in the original cannot be added in to make the story more vivid.

If a sentence requires a detail that is not present in the original text, Humanizer should ask the user rather than guess.

This is very important.

Because "de-AI flavor" can easily go to the other extreme:

To make the article sound like a real person, add some seemingly real details.

For example:

"I was standing on the street corner, drinking a cup of coffee..."

If the original author never provided this experience, then this is not a polish, but a fiction.

Humanizer's design choice was to strictly distinguish the two.

Nature does not mean fiction.

9. It also supports "imitating your writing style"

Humanizer does not require that all words end up in the same so-called "human style."

Project support provides a user's own text as a sample.

For example, you can provide two or three paragraphs of articles you have written in the past, and then ask the Humanizer:

Modify the following AI text according to my vocabulary, rhythm, punctuation and expression habits.

This makes a lot of sense.

Because "human writing" does not have a unified style.

Some people like short sentences.

Some people like long sentences.

Some people often use dashes.

Some people hardly use it.

Some people like colloquialism.

Some people like restrained, concise technical writing.

Therefore, the truly reasonable goal is not to:

"Change AI text to human text."

What should be more:

"Change the AI text to the text that this specific person will write."

This is also the interesting thing about Humanizer's style matching function.

10. A very typical example

Humanizer README gives a complete example of travel articles.

The original version is full of typical AI expressions:

"This city has completely captured my heart."

"Lisbon is a vivid testimony to the immortal spirit of Portugal."

"Every climb rewards you with amazing panoramic views."

These sentences have no grammatical problems, but lack specific personal experience.

After the revision, the article became more specific:

The hotel is located in Alfama.

I have to climb many stairs every day to buy coffee.

The Tram 28 took about forty minutes.

I ate a warm egg tart in Graça.

The author even wrote about how his legs reacted to these slopes.

Here is a very noteworthy difference:

The latter does not simply replace high-level words with low-level words.

What really changes is the structure of information.

The former is saying:

"This place is very special."

The latter is saying:

"Where I live, what I have traveled, what I have encountered, and how these experiences have made me feel specifically."

This is what Humanizer emphasizes "specific people, specific themes, specific readers."

11. How to install Humanizer?

Humanizer can be installed via Skills CLI.

For example:

npx skills add blader/humanizer --global

After installation, you can use it directly /humanizer。

For example:

/humanizer

请把下面这段文字改得更像人写的：

……

You can also use natural language directly:

Please humanize this text: [your text]

If you want to modify a file, you can also directly provide the file path, for example:

Humanize the prose in docs/launch-post.md

The project README also provides installation methods for Claude Code, Claude Desktop and other environments.

12. What scenario is it suitable for?

Humanizer is more suitable for scenarios where there is already content but feel that the text is too templated.

For example:

technology blog

AI can be responsible for completing the first draft, and then Humanizer helps remove overly uniform expressions.

Product Copywriting

It is especially suitable for changing the "functional introduction AI copy" to be more direct.

personal essay

Combining your own writing samples can make the content generated by AI closer to personal habits.

project documentation

For technical and reference texts, Humanizer does not require "writing lively", but emphasizes keeping them neutral, clear, and direct.

AI-assisted writing workflow

It can also be an additional editing stage:

AI 生成初稿
      ↓
事实检查
      ↓
Humanizer
      ↓
人工审阅
      ↓
最终稿

In particular, it should be emphasized here:Humanizer is not a fact-checking tool and should not replace manual review.

13. The true value of Humanizer

Perhaps the most noteworthy thing about Humanizer is not that it lists 25 "AI writing characteristics."

What's really interesting is the writing concept behind it:

Natural words are produced not by adding "human words", but by reducing templated decisions.

If an article uses the same structure for every paragraph, even if "delve" is replaced by "in-depth discussion", it may still have a strong AI flavor.

If each viewpoint is automatically followed by a summary sentence saying "This is not only... more...", then it will be useless to change more synonyms.

The truly effective changes are often:

• Delete unnecessary sentences;
• Delete empty emphasis;
• Reduce the mechanical platoon ratio;
• Use specific facts;
• Retain the author's own judgment;
• Let the sentence length change naturally;
• Not complete for the sake of completeness;
• Don't use complex words for "advanced" purposes.

In other words:

Humanizer is not teaching AI how to "pretend to be an adult."

It's more like reminding AI:

Less templates, more specific; less rhetoric, more information.

conclusion

The real problem facing AI writing may never be "Can AI write?"

Today's big language model can already generate articles with correct syntax, complete structure, and clear logic.

What is really difficult is:

Whether these words were written by a specific person, in a specific scene, and for a specific reader?

Humanizer has chosen a relatively pragmatic route: instead of trying to invent the so-called "perfect human style," it starts with common AI writing patterns, identifying and revising them item by item while preserving the facts, opinions, and author style of the original text as much as possible.

For those who are using AI to write, it provides not only a "touch tool", but also a way to observe AI text.

When you start paying attention to the mechanical trigrams, the "experts think" without objects, the "significance" without facts, the golden sentences at the end of each paragraph, and the sentences that exist to sound profound, you will find:

The so-called AI smell is often not mysterious. It's just choices in writing that are too regular, too safe, and too common.

Good editors often just delete these choices and leave the parts that truly contain information, opinions, and personal experience.

The project uses MIT License to open source.