by larashero3-dotcom · AI Tool · ★ 2.3k
Last updated: · Indexed by AgentSkillsHub · Auto-synced every 8h
Lieflat Less AI Tone 中文 | English 一个基于 283 万字语料统计的去 AI 味 skill 这个 skill 希望用语言学的研究方式,来量化"什么是 AI 味",并致力于去除 AI 写作的"AI 味"。该 skill 在 moxt.ai 制作。 研究以 629 篇文章(2,826,972 汉字,95,551 句,45,721 段)构建对照语料,其中 300 篇由五个主流模型在受控条件下生成,329 篇为人类作者的写作文本。检验了 26 项在公共讨论中被广泛认作"AI 文风"的候选特征,逐项计算生成侧与人类侧的频率比。 结果显示,11 项特征通过检验,有明显的区分度,其中区分力最强者为段首零回指评论(4.4 倍)。另有 15 项不具备区分力,部分项的方向与流行认知相反:人类使用比喻的频率为生成文本的 2.4 倍,正文设问句为 17 倍,句内同构排比亦高于生成文本。 研究同时发现,不同大模型的"AI 味"表现不同。例如破折号,DeepSeek 每千字 5.16 次、Claude 4.25 次,而 GPT 只有 0.11 次,极差达四十倍;提示性冒号最重的是 DeepSeek 与 Claude,问句小标题最重的是 Gemini。所谓"AI 文风"并非单一形态。 文末报告了六次因算子设计缺陷导致的测量失误及其修正过程,并讨论了单一模型样本、分母选择与语料不可核验三项方法局限。 关键词 生成文本识别;中文句法;语料库语言学;篇章衔接;对照研究 在 MoxtHub 打开去AI味.skill → 1 问题 大语言模型生成的中文文本存在某种可感知的风格特征,公共讨论中称之为"AI 味"。这一判断在写作者社群中具有高度共识,但支撑它的证据基本停留在个案观察层面。 流通中的判别清单大致包含以下条目:破折号使用过密;对举结构("不是 A 而是 B")反复出现;句长分布过于均匀;单字虚词(就、很、了)偏少;以具体物象包装抽象概念("记忆的仓库");设问后立即自答。这些条目通过社交媒体反复传播,逐步取得了近似常识的地位。 问题在于,它们从未被系统检验。清单中的条目来源各异,有的出自对早期模型输出的观察,有的来自英文语境的研究结论直接移植,有的仅是个别写作者的印象。其中哪些具有统计意义上的区分力,哪些是观察者偏差的产物,哪些方向甚至是相反的,缺乏可对照的证据。 本研究的目标不是提出新的判别清单,而是对既有清单逐项检验,并给出可复算的判定程序。 2 方法 2.1 研究设计 采用对照设计。以文本来源(模型生成 / 人类写作)为自变量,以各候选句法特征的出现频率为因变量。 有效对照要求控制三类混淆因素。 主题与文体。 生成侧的话题范围与文体设定与人类侧对齐。若两侧写作对象不属同类,观测到的差异无法归因于文本来源,可能仅反映文体差异。 模型个体差异。...
| Stars | 2,275 |
| Forks | 147 |
| Language | Python |
| Category | AI Tool |
| License | MIT |
| Quality Score | 55.0879437252621/100 |
| Open Issues | 2 |
| Last Updated | 2026-08-24 |
| Created | 2026-08-20 |
| Platforms | python |
| Est. Tokens | ~15k |
Looking for a lieflat-less-ai-tone alternative? If you're comparing lieflat-less-ai-tone with other ai tool tools, these 6 projects are the closest alternatives on Agent Skills Hub — ranked by topic overlap, star count, and community traction.
Examples of my Claude Code infrastructure with skill auto-activation, hooks, and agents
This tool has been deprecated. Use Agentic Document Extraction instead.
Visual Skills Pack for Obsidian: generate Canvas, Excalidraw, and Mermaid diagrams from text with Claude Code
All powerful skills of ClaudeKit.cc!
A tool uses Windows Filtering Platform (WFP) to block Endpoint Detection and Response (EDR) agents from report
A java agent to generate method mappings to use with the linux `perf` tool
Explore other popular ai tool tools:
lieflat-less-ai-tone is 一个基于 283 万字语料统计的去 AI 味 skill · An AI-tone removal skill grounded in a 2.83-million-character corpus study. It is categorized as a AI Tool with 2.3k GitHub stars.
lieflat-less-ai-tone is primarily written in Python.
You can find installation instructions and usage details in the lieflat-less-ai-tone GitHub repository at github.com/larashero3-dotcom/lieflat-less-ai-tone. The project has 2.3k stars and 147 forks, indicating an active community.
lieflat-less-ai-tone is released under the MIT license, making it free to use and modify according to the license terms.
The top alternatives to lieflat-less-ai-tone on Agent Skills Hub include claude-code-infrastructure-showcase, vision-agent, axton-obsidian-visual-skills. Each offers a different approach to the same problem space — compare them side-by-side by stars, quality score, and community activity.
Grades come from a rule-based scan built on the SlowMist agent-security taxonomy, covering 11 red-flag categories including credential harvesting, data exfiltration, and curl | sh installers. It is a first-layer scan, not a manual audit — we say so rather than overstate it.
The scale of the problem is documented independently: Liu et al. (2026), in a study of 31,132 agent skills, report that 26.1% contain security vulnerabilities. Our own full-catalog census is published as a citable open dataset.
Sources & who's responsible: