Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Systematically reduce AIGC detection rates in academic papers (Chinese/English). Analyzes detection reports, identifies high-impact sections, applies multi-layer rewriting strategies preserving formatting/footnotes, and verifies results. Supports 维普/知网/Turnitin platforms.
.claude/skills/telagod-reducing-aigc-detection/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 141% | 0% |
| case-12 | ✗→✓ | ▲ Improved | 198% | 0% |
| case-20 | ✗→✓ | ▲ Improved | 111% | 0% |
| case-22 | ✗→✓ | ▲ Improved | 74% | 0% |
| case-23 | ✗→✓ | ▲ Improved | 63% | 0% |
> 检测器追的是统计均匀性,反检测的本质是重新注入人类写作天然的 variance 和 imperfection。
| 场景 | 使用 | 说明 | |------|------|------| | AIGC 检测报告显示高于红线 | YES | 核心场景 | | 论文提交前预防性降 AI | YES | 不需要检测报告 | | 已有 AI 辅助写作的论文需要人性化 | YES | 最佳实践 | | 纯人工写作但误判率高 | YES | 可针对性微调 | | 想批量处理多份文件 | NO | 每篇需要定制化处理 |
| 指标 | 含义 | AI 文本特征 | 人类文本特征 | |------|------|------------|------------| | Perplexity(困惑度) | 文本可预测性 | 极低(<30) | 中高(60-120) | | Burstiness(突发性) | 句长变化幅度 | 极低,句长均匀 | 高,长短交替 | | Token 概率分布 | high-prob token 占比 | >85% | <70% |
| 平台 | 特殊机制 | 关键应对 | |------|---------|---------| | 维普 | 章节加权(摘要 1.8x,引言/结论 1.5x);拼接预警(风格断层 +10-15%) | 优先改摘要;全文风格一致 | | 知网 | 3.0+ 分析论证深度曲线;4.0 标注"结构工整度过高" | 制造浅→深螺旋节奏 | | Turnitin | 2025.8 可识别 humanizer 工具痕迹 | 不用洗稿工具,手动改写 |
优先级 = AIGC率 × 章节字数 × 平台权重系数| AIGC 率 | 策略 | 改动幅度 | |---------|------|---------| | >80% | 整段重写 | 保留核心论点,彻底换表达 | | 40-80% | 重点改写 | 换骨架、注入个人经验、打碎并列 | | 20-40% | 局部手术 | 替换 AI 特征词、打断过渡链、加短句 | | <20% | 微调或不动 | 仅修复明显 AI 模板词 |
第一层:结构层(降 60-70%,最高优先)
第二层:词汇层(降 10-15%,配合第一层)
中文 AI 高频触发词黑名单(必须替换或删除):
值得注意的是 / 综上所述 / 不可否认 / 首先…其次…最后
研究表明 / 结果显示 / 此外 / 总之 / 不仅…而且
主要体现在N个方面 / 具有重要意义 / 发挥着重要作用
在…方面 / 与此同时 / 一方面…另一方面英文 AI 高频触发词黑名单:
delve(s) / furthermore / moreover / it is important to note
comprehensive / multifaceted / nuanced / landscape / underscores
in conclusion / this report hopes to / integrates...with
The X section explains/introduces/presents/summarizes (mechanical parallelism)替换策略:不是换同义词,是换句式。「研究表明X」→ 引具体作者年份样本量。
第三层:内容注入(最难被检测)
docx 编辑策略:
| 段落类型 | 编辑方式 | 理由 | |---------|---------|------| | 无脚注/无特殊格式 | replace_full_para() — 设 Run 0 新文本,清空其余 | 安全快速 | | 含脚注引用 N] | Run 级替换 — 仅改非 superscript runs | 保留脚注 | | 含 bold/italic 段中格式 | Run 级替换或 XML 层编辑 | 保留格式标记 | | 含图表引用 | 仅改文字 runs,不动图表 XML | 防止引用断裂 |
python-docx 段落替换模板:
pythondef replace_full_para(para, new_text): """整段替换,保留首 run 格式。仅用于无脚注段落。""" if not para.runs: return para.runs[0].text = new_text for r in para.runs[1:]: r.text = ''
Run 级替换模板(保留脚注):
python# 先扫描确认哪些 runs 是脚注(superscript=True) for j, r in enumerate(para.runs): if r.font.superscript: continue # 不动 if '目标文本片段' in r.text: r.text = r.text.replace('目标文本片段', '替换文本')
改写后必须检查:
| 操作 | 为什么无效 | |------|-----------| | 仅做同义词替换 | 句式结构未变,知网 3.0 直接穿透,仅降 5-10% | | 只改标红段落不管上下文 | 风格断层触发维普拼接预警,反而加分 | | 用 humanizer/洗稿工具 | Turnitin 2025.8 已可识别工具痕迹 | | AI 改 AI 不换 prompt 策略 | 输出仍在 AI 分布内,等于原地踏步 | | 插入特殊 Unicode 字符 | 所有主流平台已修补 | | 中→英→中互翻 | 仅降 15%,且翻译腔本身可能触发 | | 越改越正式/书面 | 越规整越像 AI | | 先降查重后降 AI | 顺序错误——查重修改会引入新 AI 特征 |
| 层级 | 典型要求 | |------|---------| | C9 / 顶尖 985 | <10-15% | | 211 院校 | 15-25% | | 普通本科 | 20-30% | | 硕博论文 | <=10% |
【战场态势】各章节 AIGC 率 + 优先级排序
【改写方案】每章节策略(整段重写 / 局部手术 / 微调 / 不动)
【执行结果】改动清单 + 脚注保留确认
【预估效果】各章节预估改后 AIGC 率
【复检建议】下一步行动| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 11,601 | 16,863 | +45% | 1 | 1 | 0% | 2,115 | 5,099 | +141% | 0 | 0 | — |
case-12 | fail→pass | 8,232 | 12,171 | +48% | 1 | 1 | 0% | 1,361 | 4,060 | +198% | 0 | 0 | — |
case-02 | fail→fail | 7,188 | 15,854 | +121% | 1 | 1 | 0% | 1,071 | 4,485 | +319% | 0 | 0 | — |
case-03 | fail→fail | 8,483 | 11,894 | +40% | 1 | 1 | 0% | 1,463 | 4,159 | +184% | 0 | 0 | — |
case-04 | fail→fail | 12,223 | 16,602 | +36% | 1 | 1 | 0% | 1,745 | 4,680 | +168% | 0 | 0 | — |
case-05 | fail→fail | 16,516 | 17,654 | +7% | 1 | 1 | 0% | 2,336 | 4,832 | +107% | 0 | 0 | — |
case-06 | fail→fail | 21,214 | 27,225 | +28% | 1 | 1 | 0% | 3,004 | 6,170 | +105% | 0 | 0 | — |
case-07 | fail→fail | 6,821 | 7,685 | +13% | 1 | 1 | 0% | 1,271 | 3,512 | +176% | 0 | 0 | — |
case-08 | pass→pass | 12,358 | 13,093 | +6% | 1 | 1 | 0% | 1,814 | 4,244 | +134% | 0 | 0 | — |
case-09 | pass→pass | 13,789 | 16,916 | +23% | 1 | 1 | 0% | 2,088 | 4,690 | +125% | 0 | 0 | — |
case-10 | pass→pass | 12,536 | 13,711 | +9% | 1 | 1 | 0% | 1,949 | 4,170 | +114% | 0 | 0 | — |
case-11 | pass→pass | 11,083 | 12,459 | +12% | 1 | 1 | 0% | 1,708 | 4,037 | +136% | 0 | 0 | — |
case-13 | pass→pass | 12,985 | 10,684 | -18% | 1 | 1 | 0% | 2,204 | 3,852 | +75% | 0 | 0 | — |
case-14 | fail→fail | 20,334 | 17,069 | -16% | 1 | 1 | 0% | 2,938 | 4,804 | +64% | 0 | 0 | — |
case-15 | pass→pass | 13,490 | 15,305 | +13% | 1 | 1 | 0% | 2,136 | 4,571 | +114% | 0 | 0 | — |
case-16 | pass→pass | 18,194 | 19,454 | +7% | 1 | 1 | 0% | 2,405 | 5,060 | +110% | 0 | 0 | — |
case-17 | pass→pass | 12,993 | 14,909 | +15% | 1 | 1 | 0% | 2,050 | 4,345 | +112% | 0 | 0 | — |
case-18 | pass→pass | 12,648 | 12,831 | +1% | 1 | 1 | 0% | 1,881 | 4,139 | +120% | 0 | 0 | — |
case-19 | pass→pass | 12,185 | 16,423 | +35% | 1 | 1 | 0% | 1,795 | 4,483 | +150% | 0 | 0 | — |
case-20 | fail→pass | 15,923 | 18,899 | +19% | 1 | 1 | 0% | 2,338 | 4,923 | +111% | 0 | 0 | — |
case-21 | pass→pass | 17,338 | 9,869 | -43% | 1 | 1 | 0% | 2,541 | 3,616 | +42% | 0 | 0 | — |
case-22 | fail→pass | 11,592 | 6,786 | -41% | 1 | 1 | 0% | 1,809 | 3,142 | +74% | 0 | 0 | — |
case-23 | fail→pass | 14,815 | 9,220 | -38% | 1 | 1 | 0% | 2,148 | 3,510 | +63% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted. The headline lift of +22 percentage points is the difference between those two pass rates over the 23 comparable cases. 2 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.