Humanizer (German): Make AI Text More Human with Claude
Humanizer (German) for Claude Code and Codex: edit 72 German AI-writing patterns without Python. Open source under MIT and CC BY-SA.
AI text is often not wrong. But it sounds too smooth. Every paragraph is about the same length, every second transition says “furthermore,” and the end brings a conclusion nobody asked for.
Humanizer (German) is a free, open-source skill for Claude Code, Codex and Claude in the browser. A skill is a package of instructions that the AI loads when it needs them. This one edits German drafts: it finds typical AI-writing patterns and touches only the passages that stand out. Numbers, names and quotations stay as they are. You still need a human final check before you publish.
This article is less about how to use it and more about the reasons behind its rules. Many of them come from measurements, false alarms and a tool that was gone again after a day. Installation is further down, and you'll find the full catalog under 72 AI patterns in 10 categories.
Get started now
Open Humanizer (German) on GitHub: free, for Claude Code, Codex and Claude in the browser. The project has over 200 GitHub stars.15 The releases show the current version.
How the Humanizer Makes AI Text More Human
The Humanizer is an editing pass for existing text, not a text generator. An example:
| Before | After |
|---|---|
| “Furthermore, it is of crucial importance to implement innovative solutions seamlessly.” | “New solutions need to be introduced cleanly.” |
Here the skill spots four things: a mechanical transition (“Furthermore”), an importance filler (“of crucial importance”), AI-favorite words (“innovative solutions,” “seamlessly”) and a verb that sounds translated. In the German original that verb is the English loan implementieren, where the plain German einführen (“introduce”) is enough.
Afterward the sentence is shorter, but not smarter. That's on purpose. Where the original sentence has no real point, the skill doesn't invent one. If it finds nothing worth editing, it leaves the text alone and reports a zero edit.
You pick one of three modes: Casual for blogs, social media and newsletters, Neutral for websites, email and documentation, Formal for academic, legal and technical writing. If you don't pick one, it works in Neutral. By default you get back only the changed passages as before/after pairs, a short check and a block on the evidence. You can ask for the full text. If you hand it a file, the skill edits that file directly.
Why German AI Text Is Different
Many guides and detectors come from the English-speaking world. The Humanizer builds on two Wikipedia guides to spotting AI-generated content, one German and one English.1 The dash shows best that the German patterns are their own.
The rule about dashes first came from English sources. In August 2026 it was measured on German texts for the first time. On average, Claude uses 1.56 dashes there, GPT only 0.56. So Claude uses almost three times as many.2 GPT stands out somewhere else: its sentence lengths vary less. Two models, two fingerprints.
That did not lead to a ban on dashes. Literature, law and parliamentary speeches use dashes a lot, and quite legitimately. So the checker only fires when dashes cluster, and it knows two forms: many dashes packed into one paragraph, or a few dashes spread across many paragraphs. The second form is the more common AI habit. Both thresholds are calibrated on 39 texts written by people, without a single false alarm.3 A single dash goes unflagged.
The second German special case is quotation marks. German quotes open at the bottom („) and close at the top (“), with the mark English uses as its opener. The real signal is a mismatch: a German opening mark paired with a wrong or straight closing mark. Straight quotation marks used throughout, on the other hand, usually come from a CMS or an editor. The rule set itself shows how stubborn this error is. In the catalog, 66 German closing marks were stored as a plain straight keyboard character (ASCII, the basic character set of early computers) until an update fixed them.4
Then there are participle-I chains, the German counterpart to English -ing forms, such as gewährleistend (“ensuring”) or hervorhebend (“highlighting”). And there are false friends. A false friend is a word that looks the same in two languages but means something else: English eventually usually means “finally,” while German eventuell means “possibly.” The skill fixes meaning errors like this in every mode.
Counting Instead of Flagging Single Signals
One swallow doesn't make a summer. That's why the Humanizer doesn't treat a single signal as an AI tell, meaning a giveaway sign. It counts clusters, several signals bunched together. On their own, these stay untouched:
- correct grammar and a consistent style
- a single dash or a few typographic quotation marks
- a dry tone without specific patterns
- a single transition word such as “however” or “in addition”
- unsupported claims without other evidence or speculation patterns
How these exceptions came about is the most instructive story in the repo. A run over 21 German blog posts turned up recurring false alarms, and they became fixed exceptions.5 Schon in the sense of bereits (“already”) is not a modal particle. (Modal particles are small German words like ja, eben or schon that color the tone of a sentence without changing its content.) A capitalized Sie at the start of a sentence is often just the pronoun “she” or “they,” not the formal “you.” Straight quotation marks in HTML attributes are syntax. And a quoted AI word like nahtlos (“seamless”) is a mention, not a use.
A program that counts word forms can't tell schon the particle from schon meaning “already.” So in the end the model judges in context. Human signals stay in place: concrete details, a position that isn't clear-cut, time-bound references, genuine asides, self-corrections and varied sentence lengths.
When Humanizing Becomes a Pattern Itself
Fabricated first-person experience is the second-order tell: a sign that often appears only when someone tries to make AI text sound more human. “Last week I spoke with a client,” “Honestly,” “Don't worry”: staged anecdotes and forced chattiness are fabrication, not style. That's why the Humanizer may not invent experiences. Voice comes only from a writing sample or from facts you supply. The catalog calls it a bad reflex to add an I, we or you voice just because it sounds more human.
The same went for rhythm. The skill used to rework every third sentence opening across the board. Then came the measurement: in 20 verified human texts from before 2022, a median of 0.816 of sentences begin with the subject, about four in five. Ten humanized AI texts came in at 0.891.6 Since then the rhythm pass stops at around 0.8. Rotating further only creates the next pattern.
The rule set isn't immune either. Three catalog examples tripped a neighboring pattern themselves and had to be rebuilt.7 With rules about language, that's normal: when you break up one pattern, it's easy to build the next.
Measure First, Then Keep
Not every good idea survives the test. In August, word lists were supposed to catch advertising templates. That failed: 19 of 27 guessed candidates never showed up in real AI advertising. The successor looks for rhetorical moves instead of vocabulary, such as social proof like “more than 3,400 businesses already trust us” or stacked calls to action, and only reports something when several appear together. It has a limit too: it only recognizes phrasings it already knows. On freshly generated ads it stayed silent, because the same customer-testimonial section had a different heading every time.
Then there was a hook, a small trigger that fires after every write and reports findings to the model. In twelve comparison runs with and without the hook, the score was zero to zero. It wasn't broken: delivery was proven. It just never had anything to report. So it was dropped again in the next version, 356 lines of code and 14 tests.8 A simpler link did its job anyway.
The same honesty applies to coverage. Local scripts check only 19 of the 72 patterns with fixed rules. The model judges the other 53 in context.9
Sources Come Last
At first the skill checked style and sources in one go. That went wrong. In four texts from the effectiveness pilot, it wrote more than a hundred words each about quotation marks and passive sentences, but not a word about the planted false sources. Once source checking was separated from the style result, it flagged three of eight sources instead of none. Invented case numbers and invented people still went unmarked: they look harmless in the text and only stand out when someone looks them up.10
Since version 5.28.0 the work is fully split. The first pass stays inside the text: it marks missing footnotes, overstated claims and numbers that don't add up, but it doesn't open any source. If you want the evidence checked against the originals, you have to ask for a source sweep explicitly. It runs only on the final draft, ideally in a subagent with fresh context, meaning a second AI helper that works separately.11
The reasoning is in the docs: style and sources need two different ways of reading. You only hear rhythm when you read the text in one go. Checking sources means wandering through other people's documents. And every rewording can shift what a claim asserts, so a text that was checked before no longer counts as checked. At the end, every claim carries a status: confirmed, divergent, not checkable or missing. That includes cases where the source exists but doesn't support the claim. Without a sweep, such cases usually stay invisible. The Humanizer doesn't replace a subject-matter review.
What About AI Detectors?
Statistical detectors like GPTZero estimate two values. Perplexity measures how predictable the next word is. Burstiness measures how much sentence length and structure vary. Smooth, precise technical prose has low perplexity, and an even beat has low burstiness. The tools' labels, such as “Mechanical Precision,” only translate these two numbers. According to the catalog, they often punish exactly what makes good technical writing: technical terms, correct citations, factual clarity.12
That's why the Humanizer doesn't optimize for detector scores. Of the detector labels, it handles only two patterns: clustered colon titles and uniform sentence rhythm. The standard is editorial. Does the text sound natural to its readers? Do facts and sources stay stable? Does the form of address fit?
An optional combing gate allows one more combing pass with at most two rhythm changes. The report explicitly warns that text quality and readability can get worse. The preflight risk in the combined check says nothing about authorship either: a “low” only means that no calibrated signal fired. In advertising, social media and essays, patterns often go undetected.13
So a finding never means “AI definitely wrote this text.” It means: someone should check this passage. Translations, technical language or CMS problems leave similar traces. You'll find more on perplexity and burstiness in the experiment The AI Text Iceberg.
Install the Humanizer Skill
The base skill needs neither Python nor spaCy. What gets installed is instructions, references and optional checking scripts. These files don't start anything on their own.
Claude Code
Enter these three commands one after another in a running Claude Code session, not in a regular terminal:
/plugin marketplace add https://github.com/marmbiz/humanizer-de.git
/plugin install humanizer-de@humanizer-de
/reload-plugins
This works the same way on Windows. You don't need WSL or Python.
Codex
Run this in your terminal:
codex plugin marketplace add marmbiz/humanizer-de
Then open /plugins in Codex, select Humanizer DE, install humanizer-de and start a new session.
Claude in the Browser (claude.ai)
Without a terminal, you use the web interface. Under Settings → Capabilities, turn on “Code execution and file creation,” or the Skills section won't appear. Then go to Customize → Skills → Add → Upload a skill and upload the file humanizer-de.zip from the latest release.
Did It Work?
A skill file on disk only proves that something was copied. This test in the new or reloaded session shows whether the Humanizer is actually active:
Humanize this German text in Neutral mode:
In der heutigen dynamischen Landschaft ist es entscheidend, innovative Lösungen nahtlos zu implementieren.
The response should begin with “Less machine. More voice.”, name the mode and touch only the passages that stand out.
Want Your AI to Handle the Installation?
Paste this task into Claude Code or Codex:
Install Humanizer (German) from
https://github.com/marmbiz/humanizer-de.
Use the recommended plugin route for my platform.
Do not install add-on software or Python packages without my permission.
Then use the README test to verify that the skill is actually active,
and briefly explain what was installed.
Manual routes and updates are covered in the installation guide.
How to Use the Humanizer
After installation, a normal instruction is enough. You don't need a slash command. The better you describe the audience, text type and tone, the less the skill has to guess:
Edit this German blog post with the Humanizer in Casual mode.
Audience: people without a technical background.
Keep every fact, link, and quotation unchanged:
[paste text]
If you want, the Humanizer stores recurring style rules in a readable file called .humanizer/profile.json inside the project. It stores rules and measurement ranges, not excerpts from your text. The profile is optional, local and can be deleted at any time.
Optional checking tools
With Python 3.10 or newer, you can also measure texts locally without a language model. That makes rhythm, register and Unicode checks reproducible. The project has not measured a general quality gain from the add-on packages. Setup is described in the README.
The 72 AI Patterns in 10 Categories
Every pattern has a severity rating: HIGH (particularly critical), MEDIUM (depends on context) or LOW (noticeable only when clustered). The rating applies to the finding, not to the authorship of the whole text. The ten categories at a glance:
- Language and Tone (19): promotional language, mechanical conjunctions, participle-I chains, vague authorities like “experts say”
- Style (5): excessive bold text, emojis before headings, clustered dashes
- Communication (6): chatbot leftovers like “I hope this helps!” or “As of January 2024”
- Markup Text (6): leftover AI citation tags, broken links, invented DOIs
- Miscellaneous (3): text that breaks off, sudden style shifts
- Rhetoric and Structure (13): “In today's digital world,” rhetorical questions as fake engagement, paragraph endings with no new information
- Argumentation and Evidence (7): fabricated first-person experience, speculation that fills gaps, obscured responsibility
- Additions (4): source incongruence (the source exists but doesn't support the claim), hidden Unicode characters, Anglicisms
- Typography and Format (7): wrong quotation marks, English decimal format, English genitive apostrophe
- Title and Sentence Structure (2): colon titles, uniform sentence rhythm
The full list with severity ratings is in the pattern catalog. For a quick look at someone else's text, there's a one-page checklist.
Open Source and Free
Project code and original material are licensed under MIT. Catalog-related parts derived from Wikipedia material are licensed under CC BY-SA 4.0. Attribution and details are in the NOTICE.
The project started in January 2026 as a German adaptation of the English Humanizer by blader. It now has its own versioning scheme, German patterns, checking scripts and a test suite. The full release history is in the changelog.
Frequently Asked Questions
Can I Use It to Rewrite ChatGPT Text?
Yes. Copy the draft into Claude Code, Codex or Claude in the browser and name the audience, text type and tone. The skill isn't installed in ChatGPT itself. It edits the text created there.
Do I Need Python, and Does It Run on Windows?
Python, spaCy, LanguageTool and similar tools are optional and only meant for extra local checks. If Claude Code or Codex runs on your Windows machine, the same commands as above apply.
Does the Humanizer Also Work with Gemini?
There is no tested installation route for Gemini. The pattern catalog is available as Markdown in the repository, though, and you can use it as a reference.
Can the Humanizer Remove Watermarks from Claude or Gemini?
It doesn't promise that, and it doesn't try. Text watermarks actually in use, such as SynthID Text, work statistically through token selection, that is, through word choice, not through special characters. No character filter can find or remove them.14
Pattern 43 is plain text cleanup: it removes invisible characters like zero-width spaces because they serve no purpose in German prose. For the circulating claim that certain AI tools insert the characters U+2061 to U+2064 as watermarks, the catalog has no evidence. A match says nothing about authorship.
The Humanizer works as a style editor, and its benchmark is good German. If a watermark pushed word choice noticeably away from natural German, the Humanizer would edit those passages like any other style problem: because the text sounds off there, not because a watermark sits in it.
More from the Lab
- The AI Text Iceberg: what perplexity and burstiness reveal about a text, and what they don't.
- AI Overviews steal clicks: how GEO works: what AI answers do to content visibility.
- Control understanding with Schema and llms.txt: how LLMs read your website and what you can influence.
- Profile: who is behind the Humanizer.
Sources & References
I am aware of the irony: this article was written with AI assistance. But it was also revised with the Humanizer. A human made the final decision.