DEEP-IOS/dsh-humanizer
40Last commit Aug 15, 2026
dsh-humanizer DSH plugin
This plugin avoids common AI-removal methods that leave detectable fingerprints. Instead, it lets the model study a complete theory of human writing, then write from the perspective of a specific author, with attention to material sources, attention choices, and judgment stakes.
How to install the dsh-humanizer DSH plugin
dsh plugin --profile web add dsh-humanizerCopying does not run this command. Review the repository and version before installing the dsh-humanizer DSH plugin.
dsh-humanizer DSH plugin data source
dsh-humanizer DSH plugin snapshot date: Aug 16, 2026
discovered
What the dsh-humanizer DSH plugin can do
- Humanize writing by making the model first study the complete theory and then become the author.
- Supports both creation mode ('write a new chapter') and polish mode ('polish this text').
- Only modifies text where the reader would feel genuine discomfort, not according to external checklists.
- Preserves the original author's voice and only edits where it's not fully realized.
- Does not use any scoring, detection, or checklist post-processing.
Where the dsh-humanizer DSH plugin fits
- Writing a new chapter of a novel with a human-like narrative voice.
- Polishing a draft that feels mechanical or AI-generated without losing the original content.
- Creating content that passes as human-written in contexts where AI detection is a concern.
- Improving dialogue and narrative structure in creative writing.
Who the dsh-humanizer DSH plugin is for
- Writers and content creators who want AI-assisted writing that sounds natural.
- Developers integrating AI writing capabilities into applications that require human-like text.
dsh-humanizer DSH plugin limitations
- Does not guarantee any specific AI detection score; results vary by detector and text.
- Requires the model to read the full theory (40k characters) before each generation, which may increase latency.
- Only supports the scenarios described in the README (creation and polish); not a general-purpose rewording tool.
- Does not handle data, formulas, quotes, or sources that must not be fabricated.
dsh-humanizer DSH plugin: from the repository README
Quoted from the DEEP-IOS/dsh-humanizer README, the upstream source of the dsh-humanizer DSH plugin. Copyright remains with the original authors.
> 一个让模型在动笔之前先成为作者的 DeepSeek Harness 原生插件。 > 这里没有检测分数，没有改写清单。它做的事情，是在文字落下来之前把作者找回来。 --- ## 先说问题 市面上多数 AI 消痕工具用起来都有同样的毛病：文字越改越有 AI 味，原有排版被拆坏，句子读起来也不自然。 这个插件的第一版就犯过这个错。我们嘴上写着「配额就是新指纹」，手上却把方法做成了一张张工序卡，让用户照着执行。结果工序本身变成了新的可检测特征。所以 v0.3 干脆把方向翻了过来：模型动笔前先把整套理论完整读进脑子，在思考中成为这次要写的人，然后把理论忘掉，一口气写。 这不是换了一种提示词。往下读，你会看到这两种做法对问题的理解完全不同。 --- ## AI 味不在词上，在决定上 先看三句话。 > 这不是一次简单的失败，而是一次深刻的警醒。 > 它不仅改变了他的选择，更重塑了他的认知。 > 这意味着，他必须重新审视自己的过去。 三个句子用词不同，句式不同，读起来却像同一个人写的。把它们并排放在一起，几乎能听见同一种腔调。为什么？ 因为三个句子的底层决定是一样的。作者每次都选择同一条路：先否定一个较浅的解释，再抬到一个更抽象的高度，最后追加一层意义。词只是这条路上穿的衣服，换衣服不改变走路的姿势。 这件事可以解释一个常见的困惑：为什么删掉「此外」「值得注意的是」「综上所述」，换上一批更口语的说法，AI 味还在。因为检测器并不只看这些词。故事视镜研究用六万多个平行故事做过统计，剔除措辞、句法和韵律特征以后，单靠叙事结构特征仍然能把人类故事和生成故事分开，宏平均调和分数大约九成三。也就是说，哪怕你把句子表面的东西全换掉，只要开头的选择、情绪的载体、因果的连法、结尾的收法还是模型默认的那一套，痕迹就还在。 所以真正要动的地方不是词，是词背后的决定。谁在说，他为什么知道这些，他此刻在看什么，他凭什么下这个判断，他愿意为判断付出什么代价。这些决定对了，文字会自己带上人的位置。 --- ## 工序为什么必然失败 「把方法做成工序」这个念头很自然。理论太复杂，模型会偷懒，那就把理论拆成步骤，一步一步执行。听起来没毛病。 问题出在步骤是固定的。 假设你定了一条规矩：每章必须有一次重释、一次命名情绪、一个钩子结尾。你严格执行了二十八章。现在这二十八章有了一个共同点：它们都长着同一张脸。检测器不需要理解你的理论，它只要学会认这张脸。 更麻烦的是，这张脸是全局的。真正的写作规律往往绑在具体的材料上：这个人物受了委屈才会在对话里绕弯子，那个场景有风才值得写声音。这些规律因人而异，因文本而异，分散在成千上万个不同的决定里。而工序制造的规律不绑定任何材料，它绑定的是工序本身，所以在所有文本里长得一模一样。对分类器来说，这就是最好的特征。 第一版的十步状态
Read the full READMERepository license: MIT
dsh-humanizer DSH plugin questions
Why does the model need to read the full theory again if it already knows it?
Knowing and being present are different. The full reading brings the theory from background knowledge into the current generation context, influencing every decision. Summaries and rules only guide surface actions and cannot achieve the same depth.
How does the plugin prevent the model from skipping reading?
There is no gatekeeper for the reading step. The design relies on the completeness and persuasiveness of the reading package itself. If the model does not truly enter the author state, the resulting text will feel mechanical—that is the honest gate.
How does the polish mode ensure it doesn't destroy the original text?
The model first reconstructs the original author's voice and concerns, then only edits where the text is not fully realized—where it breaks, feels stiff, cold, empty, or says too much. Genuinely well-written sentences are preserved naturally.
Won't this method itself become a new fingerprint?
No, because it leaves no uniform shape in the text layer. It specifies whether a choice was made, not what the result looks like. Material sources, attention patterns, and judgment stakes naturally vary across contexts, so one thousand executions produce one thousand different texts.
What is the difference between this plugin and other AI text humanizers?
Most tools answer 'Where does this text look like AI? Let me fix it'—a post-processing approach. This plugin instead asks 'What does a specific author know, care about, and pay for?' before writing begins. It changes the decision layer, not just the word layer.