From org-roam-toolkit
原子笔记 (org-roam) 创建与管理规范:双语标题/标签、:ZH: drawer 折叠中文、 双层 References (inline + 汇总 + archive.today)、AI 生成内容标记、织入知识图谱原则。 当 commands 创建笔记或对话中讨论笔记格式时自动应用。 Triggers: 原子笔记, atomic note, 概念笔记, Zettelkasten, roam note, org-roam, 双语笔记, :ZH:, drawer, FILETAGS, 织入图谱, References, archive.today, ai_generated, 双向链接
How this skill is triggered — by the user, by Claude, or both
Slash command
/org-roam-toolkit:atomic-notesThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
org-roam 原子笔记的格式规范、引用规范、图谱织入原则。
org-roam 原子笔记的格式规范、引用规范、图谱织入原则。
适用范围:所有创建或修改 org-roam 节点笔记的工作。Commands /note、/study、/deep_note、/reference、/ref-extract 都按这套规范工作;自由对话中创建笔记也应遵循。
格式:English Name 中文名称 (ABBR)
规则:
示例:
Knowledge Distillation 知识蒸馏 (KD)Backpropagation 反向传播 (BP)Model Watermarking 模型水印Attention Mechanism 注意力机制Remote Procedure Call 远程过程调用 (RPC)#+FILETAGS 必须同时包含英文和中文标签。
格式:#+FILETAGS: :English_Tag:中文标签:English_Tag:中文标签:
规则:
:Intellectual_Property:IP:知识产权:示例:
#+FILETAGS: :AI_Security:AI安全:Model_Attack:模型攻击:#+FILETAGS: :Machine_Learning:机器学习:Model_Compression:模型压缩:#+FILETAGS: :Cryptography:密码学:Authentication:认证:roam_create_note 时tags 数组必须遵循同样的"成对"规则:
// ✅ 正确:每个概念双语成对
"tags": ["AI_Security", "AI安全", "Model_Attack", "模型攻击"]
// ❌ 错误:单语
"tags": ["ai-security", "model-attack"]
核心原则:英文做主线阅读流,中文用 :ZH:...:END: drawer 折叠(Emacs 中 TAB 展开)。
English one-line definition.
:ZH:
中文一句话本质定义。
:END:
** English Title 中文标题
English paragraph content.
:ZH:
中文段落内容。
:END:
** Another Title 另一个标题
- *English Term*: English description
- *English Term*: English description
:ZH:
- *中文术语*: 中文描述
- *中文术语*: 中文描述
:END:
:ZH: drawer 放中文定义** English Title 中文标题(英文在前):ZH:...:END: drawer 放中文翻译:ZH:...:END: drawer 放中文版列表笔记中的外部引用走两层结构:
正文中提到具体来源时,直接在旁边附 org-mode 链接:
Google 2024 年多次更新核心算法专门打击 AI 生成的垃圾内容
([[https://blog.google/...][Google Blog]])。
笔记末尾添加 * References 段落,汇总正文中所有外部链接,每条附 archive.today 提交链接:
* References
- [[https://blog.google/...][Google Blog: Fighting spam]] | [[https://archive.today/submit/?url=https%3A%2F%2Fblog.google%2F...][archive]]
- [[https://example.com][Source Title]] | [[https://archive.today/submit/?url=https%3A%2F%2Fexample.com][archive]]
要求:
https://archive.today/submit/?url= + URL 编码后的原始链接创建笔记后,收集所有 inline / References 中的原始 URL(非 archive.today 链接),通过 emacsclient 调用 browse-url 逐一打开 archive.today 提交页面:
emacsclient -e '(browse-url "https://archive.today/submit/?url=<URL 编码后的原始链接>")'
每条 URL 间隔 1 秒(sleep 1),避免浏览器卡顿。
所有 AI 生成的笔记 MUST 清晰标记,方便识别潜在幻觉。
:AI_Generated:AI生成:GENERATOR: AI 系统名(例 claude)MODEL: 模型版本(例 opus-4.7)GENERATED_AT: 生成时间戳(例 [2026-05-08 Wed])#+FILETAGS: :AI_Generated:AI生成:Cryptography:密码学:
:PROPERTIES:
:ID: xxxx-xxxx
:GENERATOR: claude
:MODEL: opus-4.7
:GENERATED_AT: [2026-05-08 Wed]
:END:
#+title: Knowledge Distillation 知识蒸馏 (KD)
A model compression technique where a smaller "student" model learns to mimic a larger "teacher" model.
:ZH:
通过让"学生"小模型学习"教师"大模型行为来压缩模型的技术。
:END:
/note、/deep_note、/ref-extract 创建的原子笔记:标记 AI 生成/study、/reference 创建的 reference 笔记:始终标记 AI 生成每个新建笔记必须主动找已有相关笔记建立链接。
roam_search_title / roam_search_tag 找已有相关节点roam_create_link (with bidirectional: true) 建 source ↔ target/deep_note、/ref-extract):PROPERTIES:
:ID: a1b2c3d4-e5f6-7890-abcd-ef1234567890
:GENERATOR: claude
:MODEL: opus-4.7
:GENERATED_AT: [2026-05-08 Wed]
:END:
#+FILETAGS: :AI_Generated:AI生成:Machine_Learning:机器学习:Model_Compression:模型压缩:
#+title: Knowledge Distillation 知识蒸馏 (KD)
A model compression technique where a smaller "student" model learns to mimic the soft probability distributions of a larger "teacher" model, transferring "dark knowledge" Hinton calls inter-class similarities ([[https://arxiv.org/abs/1503.02531][Hinton 2015]]).
:ZH:
通过让"学生"小模型学习"教师"大模型 softmax 概率分布的模型压缩技术。
Hinton 称类间相似性中的隐含信息为"暗知识"
([[https://arxiv.org/abs/1503.02531][Hinton 2015]])。
:END:
** Core Mechanism 核心机制
The teacher's softmax distribution contains rich inter-class relationship information.
:ZH:
Teacher 模型输出的 softmax 概率分布包含丰富的类间关系信息。
:END:
#+begin_src python
soft_target = softmax(teacher_logits / T)
loss = T**2 * KL_divergence(soft_target, soft_prediction)
#+end_src
** Related Concepts 相关概念
- [[id:<UUID-from-roam-search>][Model Compression 模型压缩]]
- [[id:<UUID-from-roam-search>][Soft Labels 软标签]]
- [[id:<UUID-from-roam-search>][Temperature Scaling 温度缩放]]
* References
- [[https://arxiv.org/abs/1503.02531][Hinton et al. 2015: Distilling the Knowledge in a Neural Network]] | [[https://archive.today/submit/?url=https%3A%2F%2Farxiv.org%2Fabs%2F1503.02531][archive]]
- [[https://en.wikipedia.org/wiki/Knowledge_distillation][Wikipedia: Knowledge distillation]] | [[https://archive.today/submit/?url=https%3A%2F%2Fen.wikipedia.org%2Fwiki%2FKnowledge_distillation][archive]]
:ZH: drawer 折叠中文(code blocks 和 References 不需要):GENERATOR:、:MODEL:、:GENERATED_AT: 三个 property + AI_Generated/AI生成 双语标签roam_search_title 查重;创建后必须按"6. 织入知识图谱"调 roam_search_title/roam_search_tag 找相关节点 + roam_create_link 建立双向链接WebSearch/fetch 验证),禁止编造npx claudepluginhub ibenjamin/org-roam-toolkit --plugin org-roam-toolkitGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.