SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillschinese-de-aigc

chinese-de-aigc

Writing & research

View repository

Updated 10-15 17:30

About this skill

Chinese Academic AIGC Reduction Skill

An AI trace eliminator for Chinese academic empirical papers. It is not "swapping synonyms," not "shuffling word order," but rathersystematically restructuring the statistical features of Chinese AI text, returning it to the language distribution of real researcher writing.

Applicable Scenarios

  • Self-checking and revising the AIGC rate before submitting to Chinese journals
  • Theses and dissertations (bachelor's/master's/doctoral) dealing with CNKI AMLC detection
  • De-AI-ification of grant applications, research proposals, and research reports
  • Scenarios where English submission has been completed and a Chinese abstract/Chinese version is needed
  • Humanization processing after a first draft is generated by tools such as CoPaper.AI

Core Philosophy

Three Wrong Approaches(many tutorials do these but they are ineffective):

  1. ❌ Synonym substitution: changing "关键" to "核心", "重要" to "显著"—detectors look at n-gram statistical distributions, not word substitutions
  2. ❌ Sentence inversion: simply changing "由于 A，所以 B" to "B，这是因为 A"—the syntactic template has not changed
  3. ❌ Handing the full-text rewrite to AI: having another AI rewrite it only swaps one kind of AI trace for another

The correct path of this Skill: targetingthe five major structural features of Chinese AIto carry out pinpoint destruction, rather than superficial modification at the word level.

The Five Major Structural Features of Chinese AI Text

Unlike English AI, the traces of Chinese large models are mainly manifested in:

  1. Excessively high density of four-character clichés: 3+ instances of "综上所述/不可否认/毋庸置疑/显而易见" and the like every 200 characters
  2. Redundancy of function words and connectives: excessive use of "然而/因此/此外/而且/并且/以及/等等"
  3. Subject avoidance and hidden passive voice: a large number of vague subjects of the "本文认为/本研究发现/相关研究表明" type
  4. Low sentence-length variance: sentence lengths cluster around 20-35 characters, lacking the rhythmic leaps of human writing
  5. Absolute conclusions: a tendency to use overly strong assertions such as "证明了/确立了/充分说明/必然导致"

Targeting the above 5 major features, this Skill provides 17 categories of fine-grained diagnostic rules(see references/patterns.md）。

Five-Step Closed-Loop Workflow

┌─────────────┐    ┌─────────────┐    ┌─────────────┐
│ 1. 定位扫描 │ →  │ 2. 诊断分类 │ →  │ 3. 差异化改写 │
└─────────────┘    └─────────────┘    └─────────────┘
                                             │
┌─────────────┐    ┌─────────────┐          │
│ 5. 二次复查 │ ←  │ 4. 五维自评 │ ←────────┘
└─────────────┘    └─────────────┘

Step 1 · Locate and Scan

Receive the text submitted by the user, scan the full text according to the 17 categories of rules in references/patterns.md ,and output a structured problem list：

## AI 痕迹定位报告

| 段落 | 原文片段 | 命中规则 | 严重度 |
|------|---------|---------|--------|
| ¶2   | "毋庸置疑，数字化转型..." | P01 四字套话 | 高 |
| ¶3   | "...此外，该研究还..."   | P04 显性连接词 | 中 |
| ¶5   | "本文认为该机制充分证明了..." | P12 绝对化断言 | 高 |

⚠️ Do notstart rewriting at this point; first let the user/author see the full picture of the problems.

Step 2 · Diagnose and Classify

Classify byparagraph function:

  • Factual statement paragraphs(methods/data/results) → low-key rewriting, preserving academic precision
  • Argumentative paragraphs(literature review/discussion/mechanism analysis) → focused rewriting, injecting the scholar's "epistemic caution"
  • Transition paragraphs(chapter transitions, paragraphs connecting the preceding and following) → major rewriting, eliminating mechanical connectives

Reference references/academic-sections.md 'schapter-by-chapter strategy tableto determine the rewriting intensity for each paragraph.

Step 3 · Differentiated Rewriting

For each item in the Step 1 checklist, fix them one by one according to the following four principles:

  1. Restore the researcher's perspective: change "this study found" into a concrete "using 2015-2023 CFPS panel data, we found" — addverifiable granularity
  2. Create sentence-length variance: within a 200-character paragraph, force the inclusion of at least 1 short sentence of ≤15 characters + 1 long sentence of ≥50 characters
  3. Eliminate explicit connectives: remove paragraph-initial connectives such as "in addition/moreover/therefore/thus", and instead usesemantic relay(the subject of the next sentence echoes the key concept in the object of the previous sentence)
  4. Absolute → cautious: downgrade all "proves/establishes/necessarily/fully demonstrates" to "provides evidence for.../is consistent with the expectations of.../may imply"

⚠️ Prohibitions：

  • It is prohibited to deliberately create typos or non-standard expressions in order to "appear human"
  • It is prohibited to use obscure words or classical Chinese-style vocabulary to forcibly reduce perplexity — this damages academic quality
  • It is prohibited to tamper with data, conclusions, or research claims — only change the expression, not the substance

Step 4 · Five-Dimensional Self-Evaluation

Perform aChinese academic version 5-dimensional scoringon the rewritten text (each dimension 1-10 points, see references/scoring.md）：

Dimension Checkpoint Weight
Specificity Whether vague expressions have been replaced with specific data/cases/authors 1.5×
Rhythm Whether sentence-length variance is ≥ 150 (mixed arrangement of 50-character long sentences + 15-character short sentences) 1.2×
Caution Whether absolute assertions have been downgraded to conditional expressions 1.3×
Implicit cohesion Whether explicit connectives (in addition/therefore, etc.) have been eliminated between paragraphs 1.0×
Researcher's tone Whether first-person research stance such as "we/our team/I" appears 1.0×

Weighted total score < 35 → return to Step 3 for another round of revision. Weighted total score ≥ 42 → pass.

Step 5 · Second Review

Re-examine the full text from a "cold reader" perspective and perform three final checks:

  1. Fluency review: Is the logic chain still coherent after rewriting? Has academic expression been harmed in order to reduce AIGC?
  2. Factual review: Have data, citations, author names, or years been mistakenly changed?
  3. Style consistency review: Is the tone consistent throughout, and is there any disconnect between "revised paragraphs vs. unrevised paragraphs"?

Output final draft + revision summary(which paragraphs were changed how much, and why).

Invocation Interface

Users can invoke it with any of the following trigger words:

  • 请对这段文本降 AIGC 检测率
  • 把这篇论文改得不像 AI 写的
  • 走 chinese-de-aigc 五步闭环
  • 诊断这段文字的 AI 痕迹，给出修改建议

Coordination with Other Skills

  • First use revision-guard to prevent excessive modification (after this Skill finishes rewriting, hand it over to it for checking)
  • Complementary to humanizer_academic: that one targets English, this one targets Chinese; a bilingual paper requires both
  • Use this Skill only after completing the first draft: do not reduce AIGC while writing, as it will disrupt your creative flow

Reference materials

  • references/patterns.md — a library of 17 types of Chinese AI-trace patterns (each type includes identification rules + typical samples)
  • references/examples.md — 12 sets of before-and-after comparisons of original text/rewrites (covering the seven main sections of empirical papers)
  • references/academic-sections.md — a table of rewriting strategies differentiated by section (abstract/introduction/literature review/methods/results/discussion/conclusion)
  • references/scoring.md — detailed rules for the five-dimension scoring scale

Important statement

The goal of this Skill isto return the text of human writing and AI-assisted writing to the language distribution of real researchers, rather than "helping AI-generated content fool detection."

  • ✅ Applicable: first drafts written by researchers themselves that are misjudged by AIGC detection as having a high AI rate
  • ✅ Applicable: mixed scenarios of AI-assisted drafting + researcher manual revision and finalization
  • ❌ Not applicable: papers fully generated by AI, hoping to pass detection with zero changes
  • ❌ Not applicable: academic misconduct scenarios, such as concealment of ghostwriting or plagiarism

Academic integrity takes priority over detection rate. No rewriting should touch research conclusions, data authenticity, or citation accuracy.