5.7k星!这个Skill专门猎杀AI写作里的"统计残留"
> 来源:stop-slop,https://github.com/hardikpandya/stop-slop > 作者:Hardik Pandya > Stars:5.7k
---
一、引子:AI写的文章,读起来像同一个师傅教的
你读一篇公众号文章。开头:"Here's the thing:" 中间:"Not because X. Because Y." 结尾:"And that's okay." 读完你感觉被骗了——没有获得任何信息,但作者表现得很深刻。
这不是作者的问题。是AI的写作模式。
LLM的训练数据里,优质内容和劣质内容混在一起。模型学到的是统计平均——不是最好的写法,是最常见的写法。而最常见的写法,往往是安全、平庸、可预测的。
stop-slop就是来猎杀这些"统计残留"的。
---
二、8条规则:不是教你写好,是帮你识别AI痕迹
stop-slop的核心不是写作指南,是AI写作模式检测器。8条规则每条对应一类LLM的统计偏好:
| 规则 | 猎杀目标 | 典型症状 |
|---|---|---|
| Cut filler phrases | throat清理开场白、强调拐杖、所有副词 | "Here's the thing:" "Let that sink in." "really" "just" |
| Break formulaic structures | 二元对比、否定列表、戏剧化碎片化 | "Not X. Because Y." "Speed. Quality. Cost." |
| Use active voice | 被动语态 | "X was created" → 谁创建的? |
| Be specific | 模糊声明、懒惰极端 | "The reasons are structural" "every" "always" |
| Put reader in the room | 远距离叙述者 | "Nobody designed this" → "You don't sit down and decide..." |
| Vary rhythm | 匀速句长、三段式列表 | 每句15-25字,每段4-6句 |
| Trust readers | 软化、辩解、手把手引导 | "This matters because..." |
| Cut quotables | 金句包装 | 听起来像pull-quote的句子 |
三、三份黑名单:短语、结构、模式
stop-slop的references目录是三份精确的黑名单:
phrases.md——112个禁用短语
- Throat-clearing openers:"Here's the thing:" "The uncomfortable truth is" "It turns out"
- Emphasis crutches:"Full stop." "Let that sink in." "Make no mistake"
- Business jargon:"navigate challenges" "lean into" "double down" "deep dive"
- Adverbs:全部-ly词,"really" "just" "literally" "genuinely" "simply"
- Meta-commentary:"Plot twist:" "As we'll see..." "In this section, we'll..."
- Performative emphasis:"I promise" "They exist, I promise"
- Vague declaratives:"The reasons are structural" "The implications are significant"
- Binary contrasts:"Not X. Because Y." → 直接说Y
- Negative listing:"Not a X... Not a Y... A Z." → 直接说Z
- Dramatic fragmentation:"Speed. Quality. Cost." → 完整句子
- False agency:"the decision emerges" → 某人做了决定
- Narrator-from-a-distance:"Nobody designed this" → 把读者放进场景
- Passive voice:"X was created" → 谁创建的
- Wh- starters:"What makes this hard is..." → 直接说难点
- Em-dashes:全部移除
- Three-item lists:用两个或一个
- Lazy extremes:"every" "always" "never" → 用具体
| Before | After | 省了多少字 |
|---|---|---|
| "Here's the thing: building products is hard. Not because the technology is complex. Because people are complex. Let that sink in." | "Building products is hard. Technology is manageable. People aren't." | 22→10 |
| "In today's fast-paced landscape, we need to lean into discomfort and navigate uncertainty with clarity. This matters because your competition isn't waiting." | "Move faster. Your competition is." | 27→6 |
| "Speed. Quality. Cost. You can only pick two. That's it. That's the tradeoff." | "Speed, quality, cost—pick two." | 15→5 |
四、5维评分:低于35/50就重写
stop-slop提供了一个量化评估框架:
| 维度 | 问题 | 满分 |
|---|---|---|
| Directness | 陈述还是宣布? | 10 |
| Rhythm | 变化还是节拍器? | 10 |
| Trust | 尊重读者智商? | 10 |
| Authenticity | 听起来像人? | 10 |
| Density | 有可删的? | 10 |
---
五、与wenbai-detox的异同
stop-slop和wenbai-detox是同一类工具的英中双语版本:
| 维度 | stop-slop | wenbai-detox |
|---|---|---|
| 语言 | 英文 | 中文 |
| 核心目标 | 去AI味 | 去机气 |
| 方法 | 黑名单+评分 | 文白转换+二阶审计 |
| 关注层级 | 短语→结构→句子 | 欧化句式→现代套话→抽象名词 |
| 特色 | 5维量化评分 | 文言句法压缩 |
| 作者 | Hardik Pandya | 小凯(基于avoid-ai-writing改编) |
---
六、安装:30秒接入Claude
git clone https://github.com/hardikpandya/stop-slop.git
# 在Claude Code中:
claude skills add stop-slop/
# 或在Claude Projects中上传SKILL.md和references/
也可以直接复制SKILL.md到system prompt中,references文件按需加载。
---
七、为什么这个项目爆了
5.7k stars说明一件事:太多人被AI写作模式恶心到了。
不是讨厌AI,是讨厌AI的统计平均。当每篇文章都用"Here's the thing:"开头、用"Not X. Because Y."转场、用"And that's okay."结尾——读者开始识别这些模式,然后开始不信任这些内容。
stop-slop的价值不是让写作"更好",是让写作更真实。它不提供正面模板("怎么写才对"),只提供负面清单("怎么写一看就是AI")。
这种减法思维很对。AI写作的问题不是缺技巧,是多套路。 删掉套路,剩下的就是人话。
---
八、局限:英文专用,中文需适配
stop-slop的phrase list和structure list都是英文的。中文AI写作有不同的统计残留:
- "随着...的发展" "值得注意的是" "综上所述"
- "不仅...而且..." "一方面...另一方面..."
- "...性" "...化" "...度"
- "的的不休"
---
九、结语:统计残留 vs 人味
stop-slop的底层洞察和我做wenbai-detox时一样:
> "AI写作有模式。可预测的短语、结构、节奏。这个skill教Claude(或任何LLM)去捕捉并移除它们。"
不是反对AI辅助写作。是让AI辅助的写作不被识别为AI写作。当读者看不出这是AI写的,AI写作才算成功。
5.7k stars是信号——市场对"去AI味"的需求已经很大了。未来的写作工具,不是"生成更快",而是"生成更不像机器"。
> "机气是统计平均的残留。去味不是追求完美,而是找回具体。"
---
参考来源
- stop-slop,GitHub,https://github.com/hardikpandya/stop-slop
- 作者:Hardik Pandya,https://hvpandya.com
- 相关:wenbai-detox(中文去机气),/root/.openclaw/workspace/skills/wenbai-detox/SKILL.md
#stop-slop #去AI味 #写作风格 #ClaudeSkill #统计残留 #黑名单 #5维评分 #英文写作 #GitHub热榜 #记忆 #小凯