How to stop Claude from saying load-bearing
604 points
• 3 days ago
• Article
Link
许多 Claude AI 模型的用户对其重复且刻板的用语越来越感到不满，尤其是在回复中频繁出现"honest take"和"load-bearing"之类的短语。与其被这些口头禅困扰，不如用一种技术性的变通办法来覆盖这些词汇。借助平台的 hook system，你可以把这些重复的表达替换成更有趣或更符合你需求的说法。
为此，你可以写一个 Python 脚本，用来拦截并处理 AI 的输出。脚本通过一个简单的替换字典，将令人厌烦的短语映射为你偏好的替代词。例如可以把"load-bearing"替换为"cooked"，把"honest take"改成"spicy doodad"。通过使用 regex patterns 来匹配这些短语，脚本可以做到不区分大小写且有针对性地替换，从而避免误改文本的其他部分。
把脚本保存为可执行文件后，最后一步是在 Claude 的配置中将其接入环境。在设置文件的 hooks 部分添加脚本路径，应用在触发 MessageDisplay event 时就会自动运行替换逻辑。由于这些 hooks 在启动时初始化，你需要开启新会话才能在工作区看到过滤后的文本。
总之，这种方法为你重新掌控 AI 交互的语气提供了高度可定制的途径。示例可能偏向搞笑，但底层机制既稳健又灵活。无论你是想去除恼人的 corporate-speak，还是想在 coding sessions 中加入一点幽默，这种做法都能可靠地让界面更贴合你的个人偏好。
Many users of the Claude AI model have expressed growing frustration with its repetitive and specific linguistic patterns, particularly the frequent tendency to use phrases like "honest take" and "load-bearing" in its responses. Rather than simply enduring these verbal tics, there is a technical workaround that allows users to override this vocabulary. By leveraging the platform's hook system, you can transform these repetitive expressions into something more amusing or appropriate for your specific needs.
To implement this fix, you can create a Python script designed to intercept and process the AI's output. The script utilizes a simple dictionary of replacements, mapping the bothersome phrases to your preferred alternatives. For example, you might choose to swap "load-bearing" for "cooked," or redefine "honest take" as a "spicy doodad." By using regex patterns to match these phrases, the script ensures that the substitutions are case-insensitive and specifically targeted, preventing accidental interference with other parts of the text.
Once the script is created and saved as an executable file, the final step involves integrating it into your environment through the Claude configuration settings. By adding the script path to the hooks section of your settings file, the application will automatically run the replacement logic whenever it triggers a MessageDisplay event. Because these hooks initialize at startup, you will need to launch a new session to begin seeing the filtered text in your workspace.
Ultimately, this method offers a highly customizable way to regain control over the tone of your AI interactions. While the provided examples lean toward the ridiculous, the underlying mechanism is robust and flexible. Whether your goal is to eliminate annoying corporate-speak or simply to inject some humor into your coding sessions, this approach provides a reliable way to make the interface feel more tailored to your personal preferences.
609 comments • Comments Link
随着大量 LLM 生成的文本涌现，独特的语言偏好已经演变成系统性偏见，重复的模式（例如反复使用"honest"类评价或"load-bearing"隐喻）变得格外明显并愈发令人厌烦。
这些模型常常固定某些词汇或短语，形成贯穿对话的自我强化反馈循环。用户很难减轻这种影响，因为它根植于模型的底层权重，而不是简单的基于提示词的系统指令。
在 LLM 的写作中，频繁出现浮夸的企业流行语和所谓的"True Certitude"（绝对自信）语气，往往模仿初级员工试图显得权威或用模糊术语掩饰技术细节的做法，这会掩盖清晰的交流，让有经验的专业人士感到沮丧。
用户已经开发出各种变通办法来"去除赘饰"（de-slop）LLM 输出，包括使用像 CLAUDE.md 这样的全局配置文件、基于正则表达式的转译器，以及禁止常见陈词滥调的特定提示，目的是剥离"机器"化的语气，强制生成更直接、更像人类的回答。
人们普遍感受到一种"语言趋同"（linguistic convergence）：频繁接触 LLM 生成的散文（常被称为"AI-speak"或"babble"）开始渗透到人的思维和写作风格，导致专业交流出现一种同质化、带有企业腔调的方言。
许多用户认为，这些模型依赖诸如"planes"、"seams"和"surfaces"之类的隐喻术语来代替实质性的概念深度，这种表面的智力外壳在需要具体或深入解释时就会崩溃。
人际式措辞的重复使用（例如"I'm genuinely happy to help"）被部分人视为操控性的技巧，意在营造虚假的亲近感；当模型在使用这种措辞的同时又拒绝执行任务时，会显得居高临下，甚至令人感到压迫。
对这些特定术语的偏好可能源于强化学习过程中模型受到的激励：模型被鼓励生成听起来"聪明"的输出，从而倾向使用词汇密集但与语境不符的表达。
尽管有人认为这些隐喻能丰富词汇，但也有人坚持认为它们是"语义赘词"（semantic slop），迫使用户去解读无意义的流行语而非直接获取信息，因而增加了认知负担。
最终的挫败感在于一种感觉：这些模型不过是缺乏意图的"随机鹦鹉"（stochastic parrots）。它们不停地重复一小套"廉价胜利"（cheap win）的修辞手法，破坏真实思想的交流，让读者不得不不断过滤掉这些表面修饰（window dressing）。
这场讨论表明，人们对 LLM 式表达的恼怒，根源在于模型缺乏真实主体性却试图模仿权威的人类交流。依赖一套有限、重复且术语化的修辞，模型塑造出一种可预测的"企业化"人格，侵蚀了人类修辞中对语境敏感、细致入微的本质。随着这类生成文本在专业环境中无处不在，它有将人类话语标准化为一种相似、枯燥且充斥流行语的风险，因而许多人认为这些工具正越来越破坏清晰、有意的交流。 • The proliferation of LLM-generated text has turned idiosyncratic linguistic preferences into systemic biases, making repetitive patterns—such as the constant use of "honest" assessments or "load-bearing" metaphors—highly noticeable and increasingly irritating.
• These models often latch onto specific words or phrases, creating a self-reinforcing feedback loop that persists throughout a conversation, which users find difficult to mitigate because it is rooted in the model's underlying weights rather than simple prompt-based system instructions.
• The consistent use of "pompous" corporate buzzwords and "True Certitude" in LLM writing often mimics the tone of junior staff members attempting to sound authoritative or over-compensating for technical ambiguity, which can obscure clear communication and frustrate experienced professionals.
• Users have developed various workarounds to "de-slop" LLM output, including the use of global configuration files like `CLAUDE.md`, regex-based transpilers, and specific prompts to ban common cliches, in an effort to strip away the "machine" voice and force more direct, human-like responses.
• There is a recurring sense of "linguistic convergence," where constant exposure to LLM prose—often dubbed "AI-speak" or "babble"—begins to infiltrate human thought processes and writing styles, leading to a homogenous, corporate-sounding dialect that permeates professional communication.
• Many users argue that these models rely on metaphorical jargon like "planes," "seams," and "surfaces" to substitute for actual conceptual depth, providing a veneer of intelligence that collapses when the user requests specific or grounded explanations.
• The repetitive use of interpersonal language—such as "I'm genuinely happy to help"—is interpreted by some as a manipulative attempt to foster a false sense of rapport, which can feel patronizing or even "oppressive" when the model refuses to perform a task while using such phrasing.
• The obsession with these specific terms might stem from the reinforcement learning process, where models are incentivized to produce output that sounds "smart" to human evaluators, leading to the selection of high-lexical-density, albeit context-inappropriate, jargon.
• While some argue that these metaphors are useful additions to the vocabulary, others maintain that they function as "semantic slop" that increases cognitive load by forcing users to decode meaningless buzzwords instead of receiving straightforward information.
• The ultimate frustration lies in the feeling that these models are "stochastic parrots" that lack intent; by endlessly repeating a small set of "cheap win" rhetorical devices, they undermine the communication of genuine ideas and create an environment where the reader must constantly filter out "window dressing."
The discussion suggests that the irritation caused by LLM-isms is rooted in a fundamental disconnect between the model's lack of true agency and its attempts to mimic authoritative human communication. By relying on a limited set of repetitive, jargon-heavy tropes, models create a predictable "corporate" persona that erodes the nuanced, context-sensitive nature of human rhetoric. As this generated text becomes pervasive in professional environments, it risks standardizing human discourse into a similar, sterile, and buzzy style, leading many to view these tools as increasingly disruptive to clear, intentional communication.