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 a similar 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 and Codex. It finds typical German AI-writing patterns and edits the passages that stand out. Its workflow protects and checks numbers, names, sources, and quotations; human review is still required before publication. You can also hand a ChatGPT draft to Claude or Codex and have it revised there.
The base skill works without Python, spaCy, or other add-on software. Install the skill, specify the type of text and tone, then review the result. Below you will also find the complete catalog of 72 AI patterns in 10 categories.
Get started now
Open Humanizer (German) on GitHub — free and available for Claude Code and Codex. The public version checked for this article is v5.19.0 from August 14, 2026.
If skills are new to you, jump straight to installation. It includes the recommended route, a Windows note, and a prompt that lets your AI handle the installation for you.
How the Humanizer Makes AI Text More Human
The skill does more than swap words. It reviews language, structure, rhythm, and evidence in context. It does not replace external research or subject-matter review.
| Before | After |
|---|---|
| “Furthermore, it is of crucial importance to implement innovative solutions seamlessly.” | “We also need to introduce new solutions smoothly.” |
The example becomes shorter and more direct without adding a new claim. That boundary matters: the Humanizer should edit a usable draft, not invent missing substance.
It works with 72 documented German patterns. They include mechanical transitions, monotonous sentence sequences, vague authorities, false source formats, and fabricated first-person experiences. A counter-check protects passages that are already clean. If your text needs only two small edits, it receives only two small edits.
Three modes adapt the edit to the text:
| Mode | Suitable for | Editing approach |
|---|---|---|
| Casual | Blog posts, social media, newsletters | more personality and a livelier rhythm |
| Neutral | Emails, websites, product copy | remove AI patterns while keeping a neutral tone |
| Formal | Academic, legal, and technical writing | edit conservatively and protect the structure |
If you do not specify a mode, the skill uses Neutral. Casual is often the better starting point for a blog post.
You typically receive:
- the selected mode
- the main patterns found
- the edited text
- a short check of tone, quality, and factual stability
When you provide a file and request changes, the skill edits the file directly and briefly summarizes what changed.
Install the Humanizer Skill
The base skill requires neither Python nor spaCy. Installation adds instructions, references, and optional checking scripts. These files do not run by themselves and do not install other software.
My recommendation as the author
The Humanizer works without Python. For the best results, I still recommend the full setup with the optional checking tools. They make rhythm, register, Unicode, and edge-case checks more precise and reproducible. In this configuration, Humanizer (German) is the best free tool for German AI text that I know.
Claude Code
Enter these three commands one after another in a running Claude Code session, not in a regular terminal:
/plugin marketplace add marmbiz/humanizer-de
/plugin install humanizer-de@humanizer-de
/reload-pluginsThe first command adds the marketplace, the second installs the Humanizer, and the third reloads the plugin in the current session. Alternatively, restart Claude Code.
This works the same way on Windows. The commands run in Claude Code; you do not need to configure a Python path or install WSL.
If the first command stops with an access message such as Permission denied (publickey), try the full HTTPS address:
/plugin marketplace add https://github.com/marmbiz/humanizer-de.gitCodex
Run this in your terminal:
codex plugin marketplace add marmbiz/humanizer-deThen open /plugins in Codex, select the Humanizer DE marketplace, and install humanizer-de. Start a new session afterward; the bundled skills become available only there.
Depending on the Codex interface, you may also find the plugin in the Plugin Directory or workspace settings. If it does not appear, check the workspace, availability, and permissions, then refresh the view or restart Codex. The direct CLI route codex plugin add humanizer-de@humanizer-de is also available.
Did It Work?
A skill file on disk proves only 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 described in the README on GitHub. The plugin is simpler for getting started.
How to Use the Humanizer
After installation, a normal instruction is enough. You do not need a special slash command:
Humanize this German text in Neutral mode:
[paste text]For a blog post, you can provide more context:
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]The same workflow makes ChatGPT drafts more human: copy the draft into Claude Code or Codex and specify the audience, text type, and desired tone. The better these three details are, the less the skill has to guess.
Your Rules, Never Your Texts
If you choose, the Humanizer stores recurring style rules in a readable file named .humanizer/profile.json inside the project. It might record that your sentences should stay short or that certain phrases should be avoided.
The profile stores rules and measurement ranges, not excerpts from your text. It is optional, local, and can be deleted at any time. You can disable it for a run without personal preferences.
Why German AI Text Is Different
Many AI detector tools and writing guides focus on English. German texts have their own stumbling blocks.
For example:
- Participle-I constructions such as “gewährleistend” or “hervorhebend” quickly sound translated in German.
- Clusters of dashes often do work that a period, comma, or colon would handle better.
- One “darüber hinaus” is unremarkable. Repeated at the start of paragraphs, it sounds mechanical.
The German Wikipedia has produced its own documentation. The English Wikipedia has a comparable page. The Humanizer uses both as foundations.
The 72 AI Patterns in 10 Categories
Every pattern has a severity rating: HIGH (particularly critical), MEDIUM (context-dependent), or LOW (noticeable mainly when clustered). The rating applies to the finding, not the authorship of the entire text.
1. Language and Tone (19 Patterns)
| Pattern | Severity | Example |
|---|---|---|
| Symbolic overload | HIGH | “stands as a testament to” |
| Promotional language | HIGH | “breathtaking,” “unique” |
| Meta commentary | HIGH | “It is important to note” |
| Mechanical conjunctions | HIGH | “Furthermore,” “Moreover” |
| Section summaries | HIGH | “Overall,” “In summary” |
| Participle-I constructions | HIGH | “gewährleistend,” “hervorhebend” |
| Vague authorities | HIGH | “Experts say,” “Studies show” |
| Misplaced “Conclusion” | MEDIUM | a conclusion heading where none belongs |
| Too-perfect conclusions | MEDIUM | “Despite X, Y faces Z” |
| Negative parallelisms | MEDIUM | “not only... but also,” clipped fragments such as “No guessing.”, “No compromises.” |
| Tricolon overuse | MEDIUM | groups of three for no real reason |
| False ranges | MEDIUM | “from traditional to modern” |
| Abstract-noun stacking | MEDIUM | “various measures,” “central aspects” instead of the concrete thing |
| Synonym rotation | MEDIUM | “the Hanseatic city,” “the city on the Elbe” for the same referent |
| Modal-particle anomaly | LOW | conversational German with no “ja,” “eben,” or “wohl” at all, or far too many |
| AI-marker vocabulary | MEDIUM | clusters of “illuminate,” “delve,” “exciting,” “the digital landscape” |
| Copula avoidance | MEDIUM | “serves as,” “possesses,” “represents” instead of plain “is” or “has” |
| Fake-analysis appendix | MEDIUM | “...which underscores/proves/demonstrates X,” a relative clause that adds no information |
| Comparative framing | MEDIUM | “less X than Y,” a comparison template instead of a direct, testable description |
Before (AI):
The breathtaking city with its rich cultural heritage stands as a testament to the artistic brilliance of past generations.
After (human):
The city has a long history. Its monuments show the craftsmanship of the Middle Ages.
The difference: fewer adjectives, more substance.
2. Style (5 Patterns)
| Pattern | Severity | Why It Signals AI Writing |
|---|---|---|
| Excessive bold text | MEDIUM | AI makes everything important |
| False lists | LOW | bullet points where none belong |
| Emojis before headings | LOW | 🎯 Goals, 📊 Data, 🚀 Launch |
| Dash overuse | MEDIUM | replacement hierarchy: period > comma > colon > semicolon > parentheses > rewrite; detects paired asides and dash variants (–, —, --) |
| Structural register collapse | MEDIUM | informal markers over fully formed written prose; sentence and paragraph architecture reveal it, not the vocabulary |
3. Communication (6 Patterns)
Chatbot remnants have no place in professional writing:
| Pattern | Severity | Example |
|---|---|---|
| Letter-like writing | HIGH | “Subject:”, “Kind regards” |
| Collaborative communication | HIGH | “I hope this helps!” |
| Knowledge-limit references | HIGH | “As of January 2024...” |
| Prompt refusal | HIGH | “As an AI model, I cannot...” |
| Placeholder text | HIGH | “[insert name]” |
| Links to search queries | HIGH | Google searches instead of real sources |
4. Markup Text (6 Patterns)
Technical mistakes AI makes:
- Markdown instead of wikitext (MEDIUM)
- Broken wikitext and AI artifacts such as oaicite tags, contentReference spans, and turn0search0 references (MEDIUM)
- Broken links (MEDIUM)
- Citation fabrication: invented DOIs, hallucinated publications, nonexistent journals, and utm_source parameters (HIGH)
- Incorrect reference formats (MEDIUM)
- Wrong categories (MEDIUM)
5. Miscellaneous (3 Patterns)
- Abrupt cutoffs, where the text ends in the middle of a sentence (LOW)
- Style shifts: suddenly formal, then casual again (MEDIUM)
- First person in metadata: “I improved this article” (LOW)
6. Rhetoric and Structure (13 Patterns)
| Pattern | Severity | Example |
|---|---|---|
| Persuasive authority phrases | MEDIUM | “At its core,” “In reality,” “The real question is” |
| Signposting | MEDIUM | “Let's take a look,” “Here is what you need to know” |
| Fragmented headings | LOW | generic one-liner immediately after a heading |
| Rhetorical questions as fake engagement | MEDIUM | “But what does this mean?”, “Have you ever wondered?” |
| Universal-human-experience opener | MEDIUM | “Since time immemorial,” “Since the dawn of civilization” |
| “In today's X world” framing | MEDIUM | “In today's digital world,” “In the age of...” |
| Aspirational corporate closing | MEDIUM | “well positioned,” “the possibilities are endless” |
| Diff-anchored writing | MEDIUM | “has now been added” when the text should describe the current state |
| Aphorism formulas | MEDIUM | “X is the language of Y,” “X becomes a trap”; a pleasant-sounding empty formula instead of a concrete claim |
| Isometric document | MEDIUM | every paragraph 3–5 sentences, every section the same length, every aspect equally weighted |
| Markerless closure compulsion | MEDIUM | every paragraph ends with an evaluative wrap-up that adds nothing (“This lays the foundation.”) |
| Announcing cleft sentence | MEDIUM | “What surprised me was...” stages a plain statement as a moment of insight instead of making it |
| Retroactive pseudo-nuance | MEDIUM | “More precisely...” followed only by a softer repetition, with no condition, exception, or mechanism |
Before (AI):
In today's digital world, a strong online presence is essential for businesses. But what does that mean in practice? With this strategy, the company is well positioned for the future.
After (human):
Companies that do not appear on Google lose customers to visible competitors. A maintained profile and a fast website win most of them back.
7. Argumentation and Evidence (7 Patterns)
| Pattern | Severity | Example |
|---|---|---|
| Passive constructions and subjectless fragments | MEDIUM | “was carried out,” “No configuration required.” instead of an active sentence |
| Conditional stacking | MEDIUM | clusters of “if/in case/provided that” clauses in conclusions instead of a direct statement |
| Miscalibrated epistemic confidence | MEDIUM | swings between “fundamentally changed” and “might perhaps” |
| Speculative gap-filling | HIGH | “keeps a low profile,” “probably,” even though the source is missing |
| Fabricated first-person experience | HIGH | “When I spoke to a client last week...” without a real source for the anecdote |
| Obscured responsibility | MEDIUM | “The strategy decided”; an abstract subject performs an attributable action and hides the actor |
| Pseudo-therapeutic validation | HIGH | “You are not too sensitive”; an unsupported diagnosis of the reader |
LLMs like to hide the actor behind passive voice and subjectless sentences. At the same time, they stack conditions where a direct statement would do. Most conspicuous is the shift between overstatement (“without doubt,” “revolutionary”) and excessive hedging (“seems possibly”) within a few sentences. When sources are missing, another problem appears: instead of writing “not substantiated,” AI fills the gap with plausible assumptions.
Fabricated first-person experience is the second-order tell: it often appears only when somebody tries to make AI text “more human.” Staged anecdotes and forced conversational phrases (“Honestly,” “Don't worry”) are fabrication, not style. The Humanizer therefore never invents experiences while rewriting. Voice comes only from the writing sample or explicitly supplied facts.
8. Additions (4 Patterns)
Four further patterns come from Wikipedia's guidance on signs of AI-generated content and its quick AI check:
| Pattern | Severity | Example |
|---|---|---|
| Source incongruence | HIGH | the source exists but does not support the claim |
| Hidden Unicode characters | HIGH | zero-width space (U+200B), soft hyphen, BOM, bidi controls |
| Standard sections without substance | MEDIUM | “Future perspectives” plus unsupported filler; do not shorten it, but make it concrete or integrate it |
| Anglicism structures | MEDIUM | hard calques and false friends: “at the end of the day,” English eventually rendered as German “eventuell,” English actually rendered as German “aktuell” |
Source incongruence is particularly tricky: the source exists, the DOI is valid, and the author wrote the publication. The paper simply does not support the claim in the text. It is a classic LLM hallucination pattern that simple fact-checking tools do not catch. A typical false friend is English eventually: it usually means “finally,” while the German “eventuell” means “possibly.” The Humanizer corrects semantic mistakes like these regardless of mode.
9. Typography and Format (7 Patterns)
This category captures text that is convincing in substance but stands out through typographic Anglicisms or decorative formatting.
| Pattern | Severity | Example |
|---|---|---|
| Incorrect German quotation marks | HIGH | German opener with a U+201D or ASCII closing mark instead of U+201C |
| English title-case capitalization | MEDIUM | “The Future Of Digital Transformation” applied to a German heading |
| English decimal or date format | LOW | “3.5 percent,” “May 12, 2026” in German text |
| English genitive apostrophe | MEDIUM | “Martin's Profil” instead of “Martins Profil” |
| Bullet-point punctuation | LOW | periods on bare keywords, inconsistent lists |
| Obsessive parataxis | MEDIUM | four or more same-shaped main clauses without a subordinate clause |
| Markdown structure artifacts | MEDIUM | one-line tables, skipped heading levels (H2→H4), a --- separator immediately before a heading |
The quotation-mark problem is particularly unpleasant: Claude can choose the wrong German closing mark, and prompting alone does not reliably prevent it. The Humanizer flags those passages; a linter can check them as well. Not every odd quotation mark is a tell. Only the asymmetry is a real AI signal: a German opening quotation mark (U+201E) paired with a wrong or straight closing mark instead of the correct U+201C. Consistently straight quotation marks are usually a CMS or editor artifact, not an AI tell; consistently English quotation marks are, at best, weak evidence. Treating every straight quotation mark as AI produces false alarms.
Obsessive parataxis is subtler: every individual sentence is correct and readability is high, but the monotony gives the machine away. Only when this rhythm genuinely clusters should some sentences be joined. Deliberate staccato, for example in advertising or manifestos, should remain untouched.
Before (AI):
The Team Analyzed The Data. The results were clear. Conversion rose by 3.5 percent. The project was completed within budget.
After (human):
The team analyzed the data and reached a clear result: conversion rose by 3.5 percent, and the project remained within budget.
10. Title and Sentence Structure (2 Patterns)
These two patterns stand out almost only when clustered. They are also the only items in the catalog that overlap with measurements used by statistical detectors.
| Pattern | Severity | Example |
|---|---|---|
| Colon-title scheme | MEDIUM | repeated “Keyword: explanatory tail” in titles and subheadings |
| Uniform sentence rhythm | MEDIUM | sentences are almost all the same length and always begin with the subject |
What Should Stay Clean
The Humanizer includes a counter-check against over-editing. Not every polished or formally correct piece of text is a problem.
These signals remain untouched on their own:
- correct grammar and consistent style
- a single dash or isolated typographic quotation marks
- a dry tone without specific patterns
- one transition word such as “however” or “in addition”
- unsupported statements without additional evidence or speculation patterns
Conversely, human signals remain intact: concrete details, ambivalent positions, time-bound references, genuine asides, self-corrections, and varied sentence lengths. The skill helps improve writing without smoothing the author away.
What About AI Detectors?
Online tools scan text and return an “AI probability.” Such detectors can provide a signal, but they are not a reliable measure of text quality.
Some methods use predictability, sentence rhythm, or other linguistic features. Others use classifiers or provenance signals. The methods are not uniform and can create false positives. A text can therefore be factually sound, edited by a human, and still appear statistically conspicuous. Conversely, a good score can hide empty, generic, or false claims. A detector flag is editorial context, not proof of authorship.
The Humanizer therefore does not optimize blindly for detector scores. The better standard is editorial: does the text sound natural to its readers? Do facts and sources stay stable? Does the form of address fit? Is the voice clear? Is there real substance instead of smooth AI rhetoric?
When a detector flag points to a real writing problem, such as monotonous sentence patterns, generic transitions, or over-polished phrasing, the Humanizer can help precisely there:
- Colon-title scheme: when the H1, the caption, and several subheadings all use “Keyword: explanatory tail,” the rhythm becomes mechanical.
- Uniform sentence rhythm: when almost every sentence is the same length and begins with the subject, the text becomes monotonous.
- Abstract-noun stacking: “Various measures to improve the traffic situation” becomes stronger when the text names the actual measures.
This position is built into the workflow: a detector flag is context for the Humanizer, not an instruction. The second revision round continues only where real quality problems remain and stops when further changes would weaken the text.
What the Humanizer Is For, and What It Is Not For
The Humanizer is an editorial tool for usable AI drafts. It helps with generic phrasing, weak transitions, register problems, monotonous rhythm, and poorly supported claims.
It does not replace research and cannot rescue false statements. If a change weakens the facts, destroys the tone, or merely makes the text look different, it should not happen.
Good German text is usually more concrete than its AI draft: verbs instead of nominalizations, well-supported details instead of vague quantities, and varied sentence lengths instead of an even beat. A text becomes human through clear decisions, not through inserted mistakes.
The skill is particularly useful for:
- content creators who use AI but still want to sound like themselves
- marketing teams working on blog posts, landing pages, newsletters, or social media copy
- professional writing where numbers, sources, claims, and tone must remain stable
- editorial teams reviewing submissions for typical AI patterns
The best standard is not a detector score. What matters is whether the text works for real readers and retains its facts.
Open Source and Free
Project code and original material are licensed under MIT. Catalog-related sections derived from Wikipedia material are licensed under CC BY-SA 4.0. Attribution and details are in the NOTICE.
The project began in 2026 as a German adaptation of the English Humanizer. It now has its own versioning, German patterns, deterministic checking scripts, and a test suite.
Its foundations are the German Wikipedia analysis, the English Wikipedia analysis, and blader's English Humanizer.
The current version, technical details, and complete release history are in the GitHub repository.
Frequently Asked Questions
How Do I Make Claude Write More Naturally?
Install the skill as described above, then give Claude Code a normal instruction such as: “Humanize this German text in Neutral mode.” The Humanizer checks the draft against the 72 German patterns and edits only the passages that stand out.
Can I Use It to Rewrite ChatGPT Text?
Yes. Copy the ChatGPT draft into Claude Code or Codex and specify the audience, text type, and tone. The skill is not installed inside ChatGPT; it edits the text created there in a supported environment.
Do I Need Python, spaCy, or Other Add-On Software?
No. The base skill works without Python and installs no add-on software. Python, spaCy, LanguageTool, and similar tools are optional and intended only for additional local checks.
Does Installation Work on Windows?
Yes, if Claude Code or Codex is running on your Windows machine. The recommended plugin route uses the same commands shown above. It does not require WSL or Python.
Can My AI Install the Skill for Me?
Yes. Use the installation task above. It explicitly says that no add-on software may be installed without your permission and that the AI should verify activation afterward.
Does the Humanizer Also Work with Codex or Gemini?
Codex is supported through its own plugin. There is no tested installation route for Gemini. The pattern catalog is available as Markdown in the repository and can be used as reference material.
Can It Detect AI Text?
The Humanizer does not determine authorship and does not return a percentage probability that a text was written by AI. It finds traces: recurring phrases, monotonous sentence rhythms, typographic artifacts, vague evidence, or over-polished rhetorical endings.
A finding therefore does not mean, “AI definitely wrote this text.” It means, “This passage deserves editorial review.” Translation, technical terminology, or CMS issues can create similar traces.
Can the Humanizer Remove Watermarks from Claude or Gemini?
It makes no such promise, and it does not try. For Claude, Anthropic has not published a watermark or a detector so far (as of August 2026). What cannot be detected cannot be deliberately removed. Gemini text, by contrast, is marked by Google with SynthID, a statistical watermark in word choice; access to the matching verification service is limited.
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
- 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.
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.