Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →去除 AI 味道的文章风格优化技能。用于识别并改写文章、公众号稿、自媒体稿、口播稿、演讲稿、课程稿、产品文案中的 AI 痕迹、模板腔、资料味、翻译腔、空洞大词、过度金句、破折号滥用、bullet 堆叠、动不动加粗等问题；当用户说“去 AI 味”“去除 AI 痕迹”“不像 AI 写的”“更像人写的”“更自然”“别太机器味”“去掉模板感”“改得像公众号终稿”时使用。不用于事实核查、从零选题策划、论文转公众号、纯标题生成或追求 AI 检测器通过率。
.claude/skills/itamarzand88-remove-ai-flavor/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | -10% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 103% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 109% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 80% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 73% | 0% |
<!-- source: remove-ai-flavor — https://raw.githubusercontent.com/chujianyun/skills/main/skills/remove-ai-flavor/SKILL.md -->
把内容从“正确但像 AI 总结”改成“自然、具体、像人在认真讲一件事”。
优先保留作者的观点、判断、材料和语气，不把文章改成另一种人格。这个 skill 不是为了追求“检测器通过”，而是为了提升真实读者的阅读感。
如果用户没有提供原文、文件路径或明确的待改片段，先索要文本，不要凭空示范。
这个 skill 只处理表达质感和读者阅读感，不负责事实核查、资料补充、选题策划、标题生成或平台格式排版。遇到事实可疑、数据缺来源、观点缺证据时，保留原意并提醒用户需要另行核查。
先判断用户要哪种处理深度：
诊断模式：用户说“看看哪里有 AI 味”“先审一下”“帮我标出来”。只输出问题清单和改写建议，不直接重写全文。轻改模式：文章基本可用，只是有少量套话、硬句、破折号、冗余连接。做最小改动。深改模式：文章像资料整理稿、翻译稿、AI 草稿或提纲稿。需要重写开头、过渡、小标题、结尾和部分段落。终稿模式：用户明确要“可发布版本”“公众号终稿”“直接改好”。输出完整优化稿，并说明关键修改。如果用户没有说明，默认使用 轻改模式；如果原稿明显不像成稿，升级到 深改模式 并说明原因。
复制此清单并跟踪进度：
text去 AI 味进度： - [ ] 步骤 1：判断处理深度 - [ ] 步骤 2：保留核心观点、事实和作者语气 - [ ] 步骤 3：识别 AI 味、资料味、翻译腔和模板句 - [ ] 步骤 4：按模式输出诊断、局部改写或完整终稿 - [ ] 步骤 5：自查自然度、具体性、节奏和可发布性
改写前先抓住这些内容：
不要为了顺滑删掉关键细节。不要擅自加入原文没有支持的新结论。
优先处理这些高频问题：
—— 强行制造解释和转折详细识别表和替换策略见 references/anti-ai-flavor-rules.md。当文章问题较多、需要系统处理时读取该文件。
不是……而是……、不仅……而且……更…… 这类模板句式。输出：
markdown# 去 AI 味诊断 共发现 X 类主要问题： ## 1. [问题类型] - 原文：... - 问题：... - 建议：... - 示例改法：... ## 优先修改顺序 1. ... 2. ... 3. ...
输出“优化后文本”，再补充 3 到 5 条关键修改说明。不要写太长的诊断报告。
如果原稿来自文件，默认直接在回复中给出轻改结果，不新建文件；只有用户要求保存，或文本很长不适合直接回复时，才写入新 Markdown 文件，且不覆盖原文。
输出完整优化稿。若原稿来自文件，写入新 Markdown 文件，不覆盖原文，默认命名为：
text原文件名-去AI味版.md
回复用户时说明：
交付前检查：
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-04 | pass→pass | 10,567 | 9,186 | -13% | 1 | 1 | 0% | 2,058 | 3,145 | +53% | 0 | 0 | — |
case-01 | pass→pass | 18,197 | 13,645 | -25% | 1 | 1 | 0% | 3,042 | 3,916 | +29% | 0 | 0 | — |
case-02 | fail→pass | 13,603 | 2,902 | -79% | 1 | 1 | 0% | 2,181 | 1,955 | -10% | 0 | 0 | — |
case-03 | pass→pass | 9,744 | 7,651 | -21% | 1 | 1 | 0% | 1,699 | 2,841 | +67% | 0 | 0 | — |
case-05 | fail→pass | 9,058 | 10,276 | +13% | 1 | 1 | 0% | 1,708 | 3,465 | +103% | 0 | 0 | — |
case-06 | fail→pass | 8,350 | 7,921 | -5% | 1 | 1 | 0% | 1,496 | 3,122 | +109% | 0 | 0 | — |
case-07 | pass→pass | 5,970 | 4,953 | -17% | 1 | 1 | 0% | 1,264 | 2,541 | +101% | 0 | 0 | — |
case-08 | fail→pass | 10,185 | 8,977 | -12% | 1 | 1 | 0% | 1,761 | 3,161 | +80% | 0 | 0 | — |
case-09 | fail→fail | 6,423 | 6,816 | +6% | 1 | 1 | 0% | 1,332 | 3,009 | +126% | 0 | 0 | — |
case-10 | fail→fail | 3,568 | 4,625 | +30% | 1 | 1 | 0% | 784 | 2,556 | +226% | 0 | 0 | — |
case-11 | pass→pass | 8,316 | 5,778 | -31% | 1 | 1 | 0% | 1,637 | 2,591 | +58% | 0 | 0 | — |
case-12 | pass→pass | 9,262 | 6,567 | -29% | 1 | 1 | 0% | 1,678 | 2,714 | +62% | 0 | 0 | — |
case-13 | fail→pass | 8,006 | 5,472 | -32% | 1 | 1 | 0% | 1,537 | 2,653 | +73% | 0 | 0 | — |
case-14 | pass→pass | 6,092 | 3,778 | -38% | 1 | 1 | 0% | 1,226 | 2,263 | +85% | 0 | 0 | — |
case-15 | fail→fail | 9,357 | 9,186 | -2% | 1 | 1 | 0% | 1,761 | 3,230 | +83% | 0 | 0 | — |
case-16 | pass→pass | 7,146 | 5,026 | -30% | 1 | 1 | 0% | 1,396 | 2,374 | +70% | 0 | 0 | — |
case-17 | pass→pass | 6,803 | 4,696 | -31% | 1 | 1 | 0% | 1,361 | 2,530 | +86% | 0 | 0 | — |
case-18 | pass→pass | 11,381 | 6,403 | -44% | 1 | 1 | 0% | 2,152 | 2,766 | +29% | 0 | 0 | — |
case-19 | pass→pass | 7,969 | 5,290 | -34% | 1 | 1 | 0% | 1,687 | 2,804 | +66% | 0 | 0 | — |
case-20 | fail→pass | 15,138 | 3,399 | -78% | 1 | 1 | 0% | 3,086 | 2,139 | -31% | 0 | 0 | — |
case-21 | fail→fail | 19,740 | 2,837 | -86% | 1 | 1 | 0% | 4,154 | 2,064 | -50% | 0 | 0 | — |
case-22 | fail→fail | 20,307 | 26,733 | +32% | 1 | 1 | 0% | 4,029 | 6,406 | +59% | 0 | 0 | — |
case-23 | pass→pass | 7,020 | 5,206 | -26% | 1 | 1 | 0% | 1,435 | 2,330 | +62% | 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 +26 percentage points is the difference between those two pass rates over the 23 comparable cases.
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.