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 · License (MIT) · Third-party notices · Contributing

Why another one

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

1. 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.
2. 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  elevated
    L13: 关于它的成因，目前有几种推测：
    L23: 如果孩子的腿疼符合以下特点，生长痛的可能性较大：

Edit (another draft, CT in pregnancy):

- 今天我们就来系统地聊一聊：孕期到底能不能做CT和核磁共振？
+ 孕期到底能不能做CT和核磁共振？
- ## 一、先分清：CT和核磁有什么不同？
- **CT（计算机断层扫描）** 本质上是X线检查，……**存在电离辐射**。
- **核磁共振（MRI）** 利用强磁场和射频脉冲成像，**没有电离辐射**。
+ ## CT和核磁有什么不同
+ CT（计算机断层扫描）本质上是X线检查，……有电离辐射。核磁共振（MRI）靠强磁场和射频脉冲成像，没有电离辐射。

Every number, dose threshold and gestational window is unchanged; only the numbered heading, the idle lead-in and decorative bold are gone.

2. A failure that became a rule

In the first edit round, the ❌ before "wrong things to do" during a febrile seizure was removed as decoration:

  ## 这些错误做法千万别做  (things you must not do)
- **❌ 掐人中**           (❌ pinch the philtrum)
+ **掐人中**             ← round 1: now reads like an instruction
+ **别掐人中**           ← round 2: after the rule change ("don't pinch the philtrum")

Panicked parents skim. The independent check flagged it as a safety issue, and the hard constraints now say: removing ❌ requires a verbal negation, bold on danger signs stays, and the only catch-all "see a doctor" line is never deleted. Full audit trail in validation/rewrite/.

What is inside

SKILL.md                    hard constraints, four modes, rule table (in Chinese)
references/patterns.md      each rule with before/after examples and keep-boundaries
references/registers.md     register rules: social media, spoken scripts & dialogue, popular science, grant/official documents, academic Chinese
references/evidence.md      sources, how conflicts between projects were resolved, folk methods checked, known gaps
scripts/zh_ai_scan.py       regex scanner: 10 measured features, density per 1,000 characters vs human/AI baselines, line numbers
validation/                 validation script, Claude-generated comparison corpus, results
tests/                      regression samples

Modes: audit (diagnose only), edit (minimal changes, with a change log and "needs author input" list), write (inventory real material before drafting), and pipeline review (fixed JSON output for multi-agent pipelines).

Evidence levels

Validation

The scanner was run on HC3-Chinese (human vs GPT-3.5 answers; open_qa, medicine, psychology) and on Claude-generated text. "Flagged" means at least one measured feature exceeds twice the human baseline.

Edit-mode results, per-feature numbers and limitations are in validation/RESULTS.md. In short:

• Current Claude models almost never use stock phrases like 综上所述 or 此外. Their AI tone is structural: idle lead-in + list, ordinal headings, long 、 enumerations. GPT-3.5 is the opposite.
• Em-dash frequency changes a lot across model versions, even within one vendor.
• The human control set is Q&A text. False-positive rates on long-form human articles with headings have not been measured yet.

Install

Claude Code:

git clone https://github.com/tangwenwen-md/chinese-de-ai-writing ~/.claude/skills/chinese-de-ai-writing

Then ask in Chinese, for example 帮我去一下 AI 味 or 这段有没有 AI 味. The scanner uses only the Python 3 standard library and works on its own:

python3 scripts/zh_ai_scan.py draft.md --register 科普

--register: 通用, 社媒, 口播, 台词, 科普, 文书, 学术.

What it does not do

• It does not aim at lowering scores from AIGC detectors (CNKI, Zhuque, GPTZero) and provides no evasion techniques. Label AI-assisted content as your platform or institution requires.
• It does not judge whether medical or legal content is correct; it only guarantees edits do not change facts or claim strength.
• It does not handle English text.

Citation

See CITATION.cff.

Acknowledgements

Rules and baselines draw heavily on lieflat-less-ai-tone, op7418/Humanizer-zh, shuorenhua, qu-ai-wei, human-writing, humanizer-zh-next and speak-human-tw (all MIT), the English and Chinese Wikipedia pages on signs of AI writing, Zhu et al. (CCL 2023) and the HC3 dataset. Details in THIRD_PARTY_NOTICES.md and references/evidence.md.