Humanizer-zh: Chinese AI writing pattern detection and editing; 24 patterns, five scoring dimensions and six major AI IDEs

Patterns include inflated significance ("this symbolizes an era"), promotional language, shallow "-ing" analysis, vague attribution ("experts point out..."), three-part constructions linked by dashes, and stock challenges-and-outlook endings. These six patterns occur frequently in press releases, white papers and annual reviews.

Patterns include common AI words ("additionally," "however," "in-depth," "vitally important," "landscape" and "crucial"), avoiding "is", negative parallelism (overusing "not A, but B"), three-part constructions (rule of three), synonym cycling, and false ranges ("everything from X to Y"). These six structures are common grammatical markers in AI-generated text.

Excessive em dashes, overused boldface, inline-heading lists, misused English title case, decorative emoji, and smart curly quotes are six visual markers that Chinese readers use to recognize AI writing. Humanizer-zh removes these markers and rewrites the paragraphs with a more natural rhythm.

Detects three types of chatbot residue: collaborative phrasing ("let us..."), knowledge-cutoff disclaimers, and flattering openings ("this is a great question"). These phrases are out of place when copied into an article, but AI-generated text often retains them.

Filler phrases ("it is worth noting" and "it should be pointed out"), excessive hedging (stacking "possibly," "perhaps" and "to some extent"), and generic positive conclusions ("overall, this is a topic worth exploring in depth") add little information and appear frequently in AI-generated text. Humanizer-zh deletes these passages or compresses them into one sentence.

Run git clone https://github.com/op7418/Humanizer-zh.git ~/.claude/skills/humanizer-zh. With a manual installation, edit SKILL.md to add prohibited brand terms, tone rules and internal examples for a company style guide.

Windows users who do not want to install Node.js or Git Bash can download the zip and extract it to %USERPROFILE%\.claude\skills\humanizer-zh\. Confirm that the directory contains both SKILL.md and README.md. Paths are case-sensitive; avoid Chinese characters in paths.

Invoke /humanizer-zh 這段 AI 生成的文字… with a Chinese text passage, /humanizer-zh @draft.md with a file, or /humanizer-zh content/**/*.md across the repository. Claude scans 24 patterns, lists matches, proposes a rewrite and runs the five-dimension score in sequence. The workflow runs in the Code / Cursor / Copilot chat interface.

Directness checks whether the text gets to the point; rhythm checks sentence-length variation; trust checks whether it treats the reader as an adult; authenticity checks whether the tone resembles human writing; concision checks whether more can be cut. The total is 50 points and can be a content CI threshold. Tenten recommends 42 points as the minimum for publication.

Humanizer-zh's SKILL.md is a plain Markdown instruction specification that agents supporting skills or custom system prompts can read directly, including Cursor, GitHub Copilot, Windsurf, OpenClaw and Cline. Teams do not need separate versions for each IDE.

Have Claude generate a draft, edit it manually, then run /humanizer-zh for the final pass. The skill removes AI writing patterns from text that already meets basic quality requirements; its effectiveness is limited on a low-quality AI draft.

Return drafts below 42 points for revision; 40 to 44 points is the caution range; 45 or more points is publication-ready. Tenten uses this threshold internally for marketing materials, blogs and newsletters. A low threshold does little filtering; an excessively high one stalls editing.

Use blader/humanizer for English and op7418/Humanizer-zh for Chinese. Install both skills under the same ~/.claude/skills/ directory; Claude selects the rules by input language, so one workflow can handle both languages.

The instructions are plain Markdown. Add brand rules to a forked SKILL.md, such as banning the Chinese terms for "empowerment" and "ecosystem," or always putting client names in quotation marks. The customized company style guide gives new team members the writing standard from their first day.

The op7418 original mainly uses Simplified Chinese vocabulary. Traditional Chinese teams should use kevintsai1202/Humanizer-zh-TW to avoid unwanted conversions between regional forms of terms such as "software" and "data." Both versions share the same 24-pattern framework; they differ only in vocabulary and quotation conventions.

Three common patterns are stock challenges-and-outlook passages ("despite the challenges... looking ahead..."), vague attribution ("industry insiders point out"), and promotional rhetoric. Humanizer-zh replaces them with concrete cases, traceable quotations and verifiable data to make the material more suitable for media citation.

Brand stories often read like institutional encyclopedia entries. The authenticity and rhythm dimensions address this problem: Humanizer-zh rewrites statements such as "our company is committed to..." as first-person statements from the founder and adds concrete times, places and situations to make the brand more recognizable.

Academic texts should retain a formal register. Humanizer-zh primarily removes inflated significance, excessive hedging ("to some extent, this may indicate..."), and common AI words while preserving rigor. The rewrite makes sentences more direct, citations more specific and conclusions more verifiable.

Some social platform algorithms are sensitive to AI writing patterns; posts with stacked dashes, excessive boldface and decorative emoji have lower reach. For social content, Humanizer-zh breaks up three-part constructions, removes emoji, and replaces boldface with natural phrasing to make posts read more like human writing.

Product launch copy often substitutes promotional language for concrete descriptions, such as "a comprehensive upgrade with an unprecedented experience." Humanizer-zh rewrites it as facts, such as "the new version adds batch processing, keyboard shortcuts and offline mode; early testers report positive feedback." Put facts first; move promotional language later or remove it.

The 541 forks indicate that many teams customize the project. The 13 open issues have active discussions.

Guizang maintains CodePilot (5.5k ★, an Electron desktop app for multi-model AI agents), NanoBanana-PPT-Skills (2.4k ★, presentation generation), ai-claude-start (a multi-account Claude Code launcher), and guizang-s-prompt (a prompt library). Their continuing maintenance records help assess the skill's long-term availability.

The original English skill is by blader, a long-standing Anthropic community contributor, and is based on Wikipedia's "Signs of AI writing." The op7418 adaptation retains the original SKILL.md structure and scoring logic while translating vocabulary, examples and categories into Chinese.

A fork for Traditional Chinese communities in Taiwan, Hong Kong and overseas. Humanizer-zh is listed in six Skills directories: Learn Skills, SkillHub, FastMCP, skills.rest, AgentSkill.work and MCP Directory.