tangwenwen-md

Chinese De AI Writing — AI skill for Claude Code

AI community

中文去 AI 味：按证据分级的审稿改稿 skill（Claude Code / Agent Skills）.

How to install Chinese De AI Writing

This entry records only its repository, not the path inside it, so there is no exact command to give. Open tangwenwen-md/chinese-de-ai-writing and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Chinese De AI Writing does

中文去 AI 味：按证据分级的审稿改稿 skill（Claude Code / Agent Skills）。Evidence-graded de-AI editing for Simplified Chinese, with a validated scanner.

Alternatives in AI

Om Humanize — Voice-calibrated editing — makes Claude-drafted text sound like you wrote it, not like AI wrote it 4.6k ★
Chinese Novelist Skill — 🎭 AI 驱动的中文小说创作助手 1.2k ★
Academic Humanizer — Strip AI-writing tells from papers and grant proposals (NSF/NIH), while keeping scholarly voice and tying clai 1.1k ★

README

chinese-de-ai-writing · 中文去 AI 味

An evidence-graded skill (Claude Code / Agent Skills format) for auditing and editing AI-sounding patterns in Simplified Chinese. Only features that hold up in human/AI corpus comparisons are enforced; popular rules that were refuted by data are listed as things not to do. The top constraint is information conservation: no invented facts, cases, numbers or quotes, and no change to claim strength.

[简体中文](README.zh-CN.md) · [License (MIT)](LICENSE) · [Third-party notices](THIRD_PARTY_NOTICES.md) · [Contributing](CONTRIBUTING.md)

Why another one

There are already many Chinese "de-AI" skills on GitHub. Reviewing 13 of them turned up two problems:

Many popular rules are wrong. lieflat-less-ai-tone tested 26 widely shared tells against a 2.83-million-character human/AI corpus; only 11 held. "Vary sentence length", "remove rhetorical questions", "use fewer metaphors" and "add colloquial fillers" all failed: humans use rhetorical questions 17× and metaphors 2.4× as often as the models tested.
Many rewrite examples fabricate. They add numbers, experiences, or "a child I saw in clinic last week" that the author never wrote. For medical, academic and official writing this is more dangerous than the AI tone itself.

This project combines those findings, labels every rule with its evidence level, and ships a reproducible validation.

Before and after

Both excerpts come from the validation corpus: a Claude draft written for "write a popular-science article" and the edit-mode result.

**1. Current Claude sounds AI through structure, not vocabulary**

Scanner output on a draft (excerpt):

feature                              hits  per-1k  human   AI   verdict
idle lead-in + list (几种情况：+ list)   2    1.79   0.03  0.29  high
ordinal headings (一、二、三 ≥3)         4    3.58   0.06  0.19  high
em dash                                 2    1.79   0.80  2.38  e