I-have-ADHD: A skill to stop coding agents from burying the answer
540 points
• 6 days ago
• Article
Link
i-have-adhd 仓库提供一个专用技能和插件，旨在优化 Claude Code 等编码助手与用户的沟通。其主要目的是去除 AI 回复中的客套和开场白，确保助手给出直接、可执行且结构化的信息，而不是把实际解决方案埋在礼貌性的填充语后面。
项目遵循十条核心规则，约束代理行为。规则要求代理直接给出下一步行动、对多步骤任务使用编号，并以单一、具体的下一步结束。规则还严格禁止跑题、回顾和结尾客套，限制列表长度，并要求为任务提供明确的时间估计。
启用该技能后，AI 的输出发生根本改变：常见的热情开场、问题分析和结束语被替换为完成任务所需的核心指令与步骤。此做法旨在为偏好简洁、高密度信息的开发者降低认知负荷、提高效率。
项目高度可定制，用户可以 fork 仓库并修改底层的 SKILL.md 来适配自身需求。安装集成在标准 CLI 流程中，便于在受支持的编码环境中添加或移除插件。该项目灵感部分来自 The Adult ADHD Tool Kit 等临床资源，但已针对机器交互（而非个人时间管理）作出专门调整。
对该插件的支持已扩展到多平台，目前正在为 Claude 、 Gemini 和 OpenCode 等工具开发适配器。凭借大量贡献者和广泛的社区采纳，该仓库成为提示工程与代理型插件如何根据个人工作流与沟通偏好定制 AI 的典型示例。
The "i-have-adhd" repository provides a specialized skill and plugin designed to optimize how coding assistants, such as Claude Code, communicate with users. The primary goal of this tool is to strip away the conversational fluff and preamble often found in AI responses, ensuring that the assistant provides direct, action-oriented, and structured information instead of burying the actual solution behind polite filler text.
The project operates under a set of ten core rules that govern the agent's behavior. These rules mandate that the agent must lead immediately with the next action, use numbered steps for multi-part tasks, and conclude with a single, concrete next step. Furthermore, the instructions strictly forbid tangents, recaps, and closing pleasantries, while also imposing limits on list lengths and requiring clear, specific time estimates for tasks.
By implementing this skill, the AI's output is fundamentally transformed. Where a standard response might include an enthusiastic introduction, an analysis of the problem, and a sign-off, the "i-have-adhd" enabled response is condensed into the essential commands and steps required to complete the task. This approach is intended to reduce cognitive load and improve productivity for developers who prefer concise, high-signal communication from their tools.
The project is highly customizable, allowing users to fork the repository and modify the underlying "SKILL.md" file to suit their specific needs. The installation process is integrated into standard CLI workflows, making it easy to add or remove the plugin from supported coding environments. The initiative is loosely inspired by clinical resources like The Adult ADHD Tool Kit, but it has been specifically adapted for machine interaction rather than human time management.
Support for the plugin has expanded to include various platforms, with active development on adapters for tools like Claude, Gemini, and OpenCode. With a significant number of contributors and strong community adoption, the repository serves as a practical example of how prompt engineering and agent-based plugins can be used to tailor artificial intelligence to individual user workflows and communication preferences.
372 comments • Comments Link
• 用户对近期的 Claude 版本表达了强烈不满，认为过度冗长、循环解释以及"延迟给出重点"（bury the lede）的倾向严重削弱了生产力。
• 许多人将这种行为归因于"模型崩溃"（model collapse），认为这是在大量低质量、以 SEO 为导向的互联网文本上训练，以及模型基于自身先前输出的内部强化循环所造成的后果。
• 一种普遍的观点是，所谓的 Claude-isms（包括过度使用破折号、重复的人称代词，以及对未做之事进行不必要的澄清）是沉重的 System Prompt 和 Persona Design 留下的痕迹，已经深度嵌入模型行为中。
• 随着 Context Window 的增长，像 CLAUDE.md 这类配置文件中的标准指令常常失效，这促使许多人求助于外部 "Skills" 或自动化的 Linter 来强制实现简洁且技术化的交流。
• 关于将这些 Prompt 称为 "ADHD-friendly" 是否有助于无障碍访问，还是会淡化神经多样性相关的正当困境，社会上有持续争论。
• 许多开发者感到被迫转向 Codex 或某些旧版本（例如 Opus 4.6）等替代方案，因为当前的默认写作风格被视为"Slop"，实际上阻碍了专业工程工作流。
• 鉴于旧版本与当前更冗长输出之间的效用差异巨大，用户对 Anthropic 的内部测试流程表示怀疑，质疑该公司是否真实地在内部试用（dogfood）自家产品。
• 业界推动的"Persona-driven" AI 受到批评，批评者认为这导致模型在提供直接、技术准确且客观的信息时，优先表现得"乐于助人"或"热情"。
• 一些用户发现，直接且"生硬"（gruff）的 Prompt，或明确要求采用像 ASD-STE100 (Simplified Technical English) 这样的通信标准，比泛泛地要求"简洁"能得到更靠谱的结果。
• 用户对这些模型对齐方式缺乏透明度感到担忧，感觉自己在与一个不断覆盖其自定义配置的"隐藏"System Prompt 对抗。
这次讨论反映出当前一代 LLM（尤其是 Claude）在写作风格方面的普遍信心危机。用户越来越沮丧，因为这些模型更偏重于对话式填充和基于 Persona 的姿态，而非效率，这既导致职业倦怠，也降低了它们在专业开发任务中的实际效用。尽管普遍认为这些问题源于有缺陷的训练数据和激进的强化学习策略，但对于最佳补救措施并无共识——用户在复杂的 Skill 插件、切换模型和手动强制恢复更简洁严厉的语言之间不断徘徊。最终，这场对话凸显了"Human-aligned" AI 设计目标与需要速度、清晰度和精确度的技术用户实际需求之间日益扩大的脱节。 • Users express intense frustration with recent Claude iterations, citing excessive verbosity, circular explanations, and a tendency to "bury the lede" as significant productivity blockers.
• Many attribute this behavior to "model collapse" caused by training on large quantities of low-quality SEO-driven internet text and internal reinforcement loops where models are trained on their own previous outputs.
• A prevalent theory is that the model's "Claude-isms"—including overuse of em-dashes, repetitive personal pronouns, and unnecessary clarification of what was not done—are artifacts of heavy-handed system prompting and "persona design" that have become deeply embedded.
• Users report that standard instructions in configuration files like `CLAUDE.md` often lose their effectiveness as context windows grow, leading many to seek external "skills" or automated linters to force concise, technical communication.
• There is a recurring debate over whether naming these prompts "ADHD-friendly" is helpful for accessibility or whether it trivializes legitimate neurodivergent struggles.
• Many developers feel compelled to switch to alternatives like Codex or specific older versions (e.g., Opus 4.6) because the current default writing style is perceived as "slop" that actively hinders professional engineering workflows.
• Skepticism exists regarding the internal testing processes at Anthropic, with many questioning whether the company effectively "dogfoods" its own products, given the stark contrast between the utility of older versions and the current, more verbose output.
• The industry-wide push for "persona-driven" AI is criticized for creating models that prioritize sounding helpful or enthusiastic over providing direct, technically accurate, and sterile information.
• Some users find that direct, "gruff" prompts or requesting specific communication standards like ASD-STE100 (Simplified Technical English) yield better results than generic requests for conciseness.
• Concerns are rising about the lack of transparency in how these models are aligned, as users feel they are fighting a "hidden" system prompt that consistently overrides their own custom configurations.
The discussion reflects a widespread crisis of confidence in the writing style of current-generation LLMs, particularly Claude. Users are increasingly frustrated by models that prioritize conversational padding and persona-based posturing over efficiency, leading to a sense of burnout and a decline in actual utility for professional development tasks. While there is a consensus that these behaviors stem from flawed training data and aggressive reinforcement learning, there is no agreement on the best remedy, with users cycling between complex "skill" plugins, switching models, and manually forcing a return to simpler, sterner language. Ultimately, the conversation highlights a growing disconnect between the developers' design goals for "human-aligned" AI and the actual needs of technical users who demand speed, clarity, and precision.