humanizer-zh Skill for Claude: Edit Chinese Writing Without Changing the Facts

humanizer-zh makes Claude edit filler, repetition and template phrases out of Chinese text while keeping your facts, certainty and voice.

Installs: 56.1K; on skills.sh
Stars: 21.7K; on GitHub
Updated: Sep 2026; last commit

This skill is an editing guide for Chinese text, written in Chinese. Paste text or name a file and Claude removes empty openers, repeated sentence shapes and stock phrases, then returns the final rewrite. Its first rule is to keep every fact, number, name and level of certainty. It will not add data to make a sentence sound specific.

It checks 31 patterns in six groups, treated as clues and not a blacklist, so a paragraph with no problem can stay as it is. The README says it does not prove who wrote a text and does not promise to pass any AI detector. It edits files only when you ask.

Good for

Chinese writers and bloggers polishing drafts that read like a template
Developers cleaning Chinese READMEs, docs and product pages
Editors who want a rewrite that adds no new claims

Install humanizer-zh

Claude Code

Copy this prompt.

Install the humanizer-zh skill in Claude Code. Run this command:

npx skills add https://github.com/op7418/humanizer-zh --skill humanizer-zh

Then check that it loaded and tell me how to use it.

2
Paste it into Claude Code. It shows the change and asks before running it.

Requires Node.js.

From its listing on skills.sh. Source.

What's in the skill

When it loads

编辑中文文章、评论和文档中的空话、重复及模板化表达，让文字更自然，同时保留事实、确定程度和作者声音。适用于润色已有文本，不用于判断作者身份或保证通过 AI 检测器。

Sections in SKILL.md

编辑约束与优先级
文体与作者声音
工作流程
输出与文件保护
模式的使用方式
A. 铺垫代替陈述
1. 不是 X，而是 Y
2. 单句收尾与戏剧性碎片
3. 格言与伪深度
4. 起跑式铺垫
5. 与假想敌辩论
B. 公式化节奏
6. 强凑三段式
7. 句式开头复读
8. 破折号当万能连接
9. 限定词堆叠

Key facts

Author: op7418 (GitHub repo owner)
Repo: op7418/humanizer-zh
Installs on skills.sh: 56,132 as of 2026-11-02
GitHub stars (repo): 21,678 as of 2026-11-02
Last repo commit: 2026-09-23
License: MIT
Files: 9 files including SKILL.md (README, CHANGELOG, LICENSE, .gitignore and a tests folder with 4 files)
Works in: Claude Code (the README shows install into ~/.claude/skills and the /humanizer-zh command)
Revision in SKILL.md: 2026-09-23, based on blader/humanizer v3.0.0 and Humanizer-zh PR 39
skills.sh audits: Gen Agent Trust Hub pass, Socket pass, Snyk pass

What it includes

Four rules when they conflict
Keep information and certainty, follow your edit scope and style, match the author's voice, then fix specific expression problems.
Editing constraints, in priority order
Four steps
Read the input and note what must stay, change only real problems, compare the rewrite against the original for added or strengthened claims, then read the result through.
Workflow
Final text for pasted text, write back only on request
Code, commands, paths, URLs, IDs, YAML front matter and table data stay unchanged. Headings keep their text and level.
Write · Output and file protection
Patterns 1 to 5
'Not X but Y' contrasts, one-line closers, sayings that sound deep, run-up openers and arguing with no one.
A. Staging instead of stating
Patterns 6 to 11
Forced triples, repeated sentence openers, dashes as the universal connector, stacked qualifiers, coined compound words and passive voice with no subject.
B. Formulaic rhythm
Patterns 12 to 18
Stock AI words, inflated significance, vague connections, tacked-on grand endings, sales language, borrowed authority and avoiding plain 'is' and 'has'.
C. Inflation and borrowed authority
Patterns 19 to 21
Decorative bold, decorative headings, and quote or punctuation habits.
D. Formulaic formatting
Patterns 22 to 25
Customer-service tone, knowledge-limit disclaimers and filler guesses, a first sentence that repeats the title, and talking about the earlier draft.
E. Chat and draft leftovers
Patterns 26 to 31
Stacked '的', '进行' plus a verb, stacked passives, four-character-phrase runs, 'with the development of...' openers and stock closers.
F. Extra checks for Chinese
Compare to the source before handing back
Look for added, dropped or strengthened claims, and check agent, time, condition, scope, negation and attribution.
Pre-delivery check

From the official docs

Questions

What does humanizer-zh do?

It edits Chinese text so it reads less like a template. Claude cuts filler and repetition, keeps your facts and level of certainty, and returns the final version. The patterns are clues to check, not words to ban.

How do I install it in Claude Code?

Use the install box above. The README's recommended route is npx skills add https://github.com/op7418/Humanizer-zh.git, and it also shows cloning into ~/.claude/skills/humanizer-zh. Then type /humanizer-zh.

Does it work on English text?

Its instructions, patterns and examples are written for Chinese. For English text, see our page for the humanizer skill, which it names as a source.

Will it beat AI detectors?

No. The README says it cannot prove authorship and does not promise to pass any AI detector. It is an editing guide for clearer, less repetitive writing.

Will it change my meaning or add facts?

It is built not to. The first rule is to keep facts, numbers, names, dates and certainty, and not to turn 'may' into 'will' or 'planned' into 'done'. If detail is missing it tells you to ask for it, not invent it.

Will it edit my files?

Only if you ask. The skill writes back to a file when you request changes, and edits prose only. Say 'suggestions only, do not change the file' to get a review without edits.

Does it run scripts or need an API key?

The skill itself is instructions, with allowed tools Read, Write, Edit and AskUserQuestion. The repo has a tests folder with a Python structure check, but the SKILL.md does not tell Claude to run it. No API key is named.

How is it related to the humanizer skill?

The SKILL.md says it is a structural revision based on blader/humanizer v3.0.0 and Humanizer-zh PR 39, with stop-slop as a reference. We cover that one on the humanizer skill page.

Information

updated Nov 2, 2026

Repository: op7418/humanizer-zh

op7418/humanizer-zh skills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 3