ANTHROPC - Claude Code - Plugins & Agent Skills & Agent SDK
参考 ANTHROP\C - Claude > ANTHROPC - Claude Opus/Sonnet 5 & Fable 5 & Mythos 5 & ANTHROPC - Claude Code & Agent SDK
参考 Model Context Protocol
Plugins - Extend Claude Code with custom commands, agents, hooks, and MCP servers through the plugin system. > Plugin marketplaces - Create and manage plugin marketplaces to distribute Claude Code extensions across teams and communities.
- anthropics / claude-code - plugins 通过 /plugin 按需安装,如 feature-dev & frontend-design;
- anthropics / claude-plugins-official: Anthropic-managed directory of high quality Claude Code Plugins. > 如 Ralph Loop Plugin - Ralph is a Bash loop (Stop hook);code-simplifier;Claude Code Setup Plugin(AI 反身性);
Agent Skills & Agent Skills - Create, manage, and share Skills to extend Claude's capabilities in Claude Code. > 1016 Equipping agents for the real world with Agent Skills > Agent SKills - A simple, open format for giving agents new capabilities and expertise.
- anthropics / skills: Public repository for Skills >
/plugin marketplace add anthropics/skills添加 Marketplace 后安装 Plugin,Git 连接需通畅;> 如 Enable Auto-Update,每次启动需保持 Git 通畅使其重新拉取,而容易报错,故可手工更新;另,调试 Plugin 时暂时 Disable 其他 Plugins; - agentskills / agentskills: Specification and documentation for Agent Skills > Public repository for Agent Skills
- Codex Skills Registry - Find the right skill for your Codex CLI projects. - A catalog of installable skills that extend Codex CLI with new capabilities, so teams can ship faster without reinventing workflow automation. > 偏 Codex 的 Skills,参考 OpenAI Codex CLI/APP & AGENTS.md & Agents SDK / AgentKit;
- travisvn / awesome-claude-skills: A curated list of awesome Claude Skills, resources, and tools for customizing Claude AI workflows — particularly Claude Code
- ComposioHQ / awesome-claude-skills: A curated list of awesome Claude Skills, resources, and tools for customizing Claude AI workflows > 参考 phidata > Agno & Composio / OOMOL - open-connector;
- numman-ali / openskills: Universal skills loader for AI coding agents
- npm i -g openskills > openskills install anthropics/skills > openskills sync
任何可以用文字描述清楚的专业流程与行为都可以制作成技能(如同 Linux 万物皆文件,Agent 万事皆文本),而 Agent 时代本不应该仅靠流程图来定义与交流业务流程与行为节点。因此,较本质仍是 Tool-Use 模式的 Model Context Protocol 更高效稳定,其颗粒度层次不只是协议(如同 JSON Schema 抽象于数据,OpenAPI 抽象于 Service),而是叠加在模型能力上 Progressive Disclosure & Composability(渐进式披露与组合)的 Skill。所谓
We're "loading in" specialized knowledge to our general agents at runtime.,针对当下 AI 的控制流和数据流混在一起的死穴,可能是一个解决方案。近期 Agentic 的谱系讨论愈加清晰,在自动的预设可控的 Workflow 和自主的灵活泛化的 Agent 的频谱之间,最后总要有个平衡点,让人机协同过程中消耗掉的脑力和 Token,可以固化传递给其他智能体(或人),这也许就是 Agent 时代源代码的编程语言(后续 PTC 更夯实了这一点)。> CC 团队是真正在大模型能力基础上不断思考人(的思考表达)机(的遵循行动)协同的边界,从 MCP 到这次以 Skill 来彰显 Plugins,都是以 AI-Coding 作为第一性。也就是说,对自家大模型能力有信心,所以针对上下文工程等形式化,在高阶层面回到软工本身的编程逻辑,比如本地操作系统能力复用、充分文本/文件(含程序代码、伪代码、半结构化等广义代码)、声明/实现解耦等等。
Agent SDK overview - Build custom AI agents with the Claude Agent SDK > 支撑 GitHub Actions & GitLab CI/CD;> 20260220 如何构建 Agent 及其上下文工程(简洁入门)
1016 Equipping agents for the real world with Agent Skills & 1017 Introducing Agent Skills
1112 Improving frontend design through Skills - Best practices for building richer, more customized frontend design with Claude and Skills.
1104 Code execution with MCP: Building more efficient agents - Direct tool calls consume context for each definition and result. Agents scale better by writing code to call tools instead. Here's how it works with MCP. > Anthropic 从 MCP 回归到 Code execution 范式,即 AI-Coding (Code-Use) 的第一性。> 1124 Introducing advanced tool use on the Claude Developer Platform - Programmatic tool calling
被 Claude 吞噬的应用层:从 MCP、Skills 到 PTC 的跨维度打击 - 程序化工具调用 - Compute over Context & Coding as Reasoning(LLM“概率”与 Code“逻辑”的正交),类似于 OS 的 Shell Scripting / Kernel Execution,模型直接编写完整脚本在安全隔离的沙箱中批量执行。另参考 Pydantic AI & DSPy & AiPy - Function Calling -> MCP -> PTC 使用 JSON 去实现 Calling,而 CodeAct 及 Python-Use 则 LLM 直接生成调用对应函数的代码,本质即权衡“可靠性”vs“灵活性”。
参考 Control Theory in AI Agents: Two Theorems You're Already Using > Principle 1: build tools that cover the environment's variety & Principle 2: the LLM needs a model of the environment
SkillOpt - Executive Strategy for Self-Evolving Agent Skills. SkillOpt treats a compact natural-language skill document as the trainable state of a frozen language agent, then learns that document through rollouts, reflection, bounded edits, and held-out validation gates.
- microsoft / SkillOpt: SkillOpt: Executive Strategy for Self-Evolving Agent Skills - SkillOpt is a text-space optimizer that trains reusable natural-language skills for frozen LLM agents through trajectory-driven edits, validation-gated updates, and deployable best_skill.md artifacts.
302.AI - 302 AI Studio - Your AI Productivity Accelerator - Your cross-platform desktop AI application. Offering powerful general AI capabilities like code generation, document summarization, and intelligent Q&A to boost your productivity. > 20260112 Skills Are Just Apps for AI
Vercel - v0 & Next.js > SKILLS - THE OPEN AGENT SKILLS ECOSYSTEM
OpenCLI & CLI-Anything - 传统软件端(包括页端)的 CLI 化,是在 Agentic Agent 与 Skills 联合蚕食下,对已有固化的业务最快的变现反应:Agent Tool Call 的天花板基本被锁死了,Tool 太多了不如走 Skill+CLI,Tool 太少了又是典型的“覆盖主力工况”。
- davila7 / claude-code-templates: CLI tool for configuring and monitoring Claude Code > aitmpl: agents & commands & settings & hooks & mcps > plugins & skills,前期调试参见 ANTHROPC - Claude Code & Agent SDK;
- wshobson / agents: Intelligent automation and multi-agent orchestration for Claude Code > Claude Code Plugins: Orchestration and Automation
- obra / superpowers: Claude Code superpowers: core skills library > 参考 SuperClaude-Org / SuperClaude_Framework,1222 安装 marketplace 暂不安装相关 plugins,能力与技术路径趋同;> 20260106 卸载 SuperClaude 后安装 superpowers 相关插件;> 0517 Codex 中 Skill 安装;> 详见 SuperClaude Framework & SuperPowers
- Matt Pocock - AIHero - grill-with-docs & Garry Tan - gstack
- EVERY - Compound Engineering
202606 期间 AI-Coding & Work Agents 进化路径(“AI 时代的源代码”):SuperClaude Framework & SuperPowers & OpenCode - Zen / OMO(LazyCodex) & Crush > Matt Pocock - AIHero - grill-with-docs & Garry Tan - gstack/gbrain > EVERY - Compound Engineering
- microsoft / waza: CLI / Framework for Agent Skills - create, test, measure and improve skill quality and effectiveness > waza
- muratcankoylan / Agent-Skills-for-Context-Engineering: A comprehensive collection of Agent Skills for context engineering, multi-agent architectures, and production agent systems. Use when building, optimizing, or debugging agent systems that require effective context management.
- formulahendry / agent-skill-code-runner: An Agent Skill that enables AI agents to run code snippets in multiple programming languages. Works with GitHub Copilot, Claude Code, and other skills-compatible agents.
- ninehills / PatentWriterAgent: PatentWriterAgent Demo
- thedotmack / claude-mem: A Claude Code plugin that automatically captures everything Claude does during your coding sessions, compresses it with AI (using Claude's agent-sdk), and injects relevant context back into future sessions. > Claude-Mem
- yusufkaraaslan / Skill_Seekers: Single powerful tool to convert ANY documentation website into a Claude skill > Convert documentation websites, GitHub repositories, and PDFs into Claude AI skills with automatic conflict detection
Agents & SubAgents
- msitarzewski / agency-agents: A complete AI agency at your fingertips - From frontend wizards to Reddit community ninjas, from whimsy injectors to reality checkers. Each agent is a specialized expert with personality, processes, and proven deliverables. > 把世界上几乎所有职位都做成了 AI 员工,包括: 前端开发、UI 设计、自媒体运营、销售、市场分析师、数据工程师、法务顾问……
垂类 Skills
-
nextlevelbuilder / ui-ux-pro-max-skill: An AI SKILL that provide design intelligence for building professional UI/UX multiple platforms
-
PleasePrompto / notebooklm-skill: Use this skill to enable Claude Code to communicate directly with your Google NotebookLM notebooks. Query your uploaded documents and get source-grounded, citation-backed answers from Gemini. Features browser automation, library management, persistent authentication, and answers exclusively from your own knowledge base. > 参考 Google AI Studio & LABS.GOOGLE - Bard > Gemini & Gemma - NotebookLM - Share,本地构建隔离 Python 环境(.venv)并执行脚本(使用 Chrome 而非 Chromium),所有内容保存在本地目录中;
-
KDense - Autonomous AI Scientist From Question To Analysis - K-Dense conducts complete scientific research from hypothesis to publication with minimal human intervention > Smarter Science - Faster, More Accurate & Autonomous Research - Any Task, Any Data & Breakthrough Ready - Discover, Validate, Publish > 调试详见 KDense - Agentic Data Scientist & Scientific Agent Skills & claude-scientific-writer
-
zechenzhangAGI / AI-research-SKILLs: Comprehensive open-source library of AI research and engineering skills for any AI model. Package the skills and your claude code/codex/gemini agent will be an AI research agent with full horsepower. Maintained by Orchestra Research. > Orchestra - AI-Native Research, From Idea to Publication - Experience how AI-native research feels. Every step empowering, intelligent, and aspirational. > 参考学术 & 研发-读写系统-使用技巧 2023-2026;
-
kepano / obsidian-skills: Claude Skills for Obsidian & heyitsnoah / claudesidian: Claude Code + Obsidian Starter Kit > 参考 Obsidian(插件)使用技巧 2023-2025;
-
参考 Vercel - v0 & Next.js,vercel-labs / agent-browser & vercel-labs / agent-skills 等;
-
参考 React Native - Expo & Wasp & Remotion - remotion,配套 Skills 生成教学视频;
-
xstongxue / best-skills: 通用高质量 Skills 合集🔥
-
sivaprasadreddy / sivalabs-marketplace: Java Application Development plugins for Claude Code > Spring Boot Skills
-
veryInc / compound-engineering-plugin: Official Claude Code compound engineering plugin
-
wshuyi / x-article-publisher-skill: Claude Code skill for publishing Markdown articles to X (Twitter) Articles
-
tfriedel / claude-office-skills: Office document creation and editing skills for Claude Code - PPTX, DOCX, XLSX, and PDF workflows with automation support > 后被官方 anthropics / skills 超越;
-
Fokkyp / SoftwareCopyright-Skill: 中国软件著作权申请材料 生成器 Skills,本 Skills 通过阅读本地项目,自动生成全套 .docx 软著申请材料,全开源,无须再付费购买任何软著申请服务
-
hugohe3 / ppt-master: AI generates a real, editable PowerPoint from any document — native shapes & animations, speaker notes voiced as audio narration, and the option to follow your own .pptx template, not slide images · by Hugo He > PPT Master - AI generates natively editable PPTX from any document — real PowerPoint shapes, not images. > 从夯到拉锐评一下中文 AI 博主的 PPT skill
-
zarazhangrui / frontend-slides: Create beautiful slides on the web using a coding agent's frontend skills
-
op7418 / NanoBanana-PPT-Skills: NanoBanana PPT Skills 基于 AI 自动生成高质量 PPT 图片和视频的强大工具,支持智能转场和交互式播放 > Pyhton 脚本调用 Google 的 NanoBanana 以及可灵的视频生成;将生成的视频和图片路径放到演示网页的代码中生成一个演示网页;同时调用本地的 ffmpeg 将图片和视频剪辑成为一个完整的演示视频;
-
op7418 / guizang-ppt-skill: AI-agent Skill for generating polished HTML slide decks: editorial magazine and Swiss layouts, image prompts, social covers, and a WebGL/low-power presentation runtime.
歸藏(guizang.ai)@op7418: 压进我十年设计经验的 PPT Skills,迎来大波更新 & 万字长文:做了些爆款 Skills 以后,我对 Skills 的看法:好 Skill 的架构:中心短,辐射厚;Skill 质量要像代码质量一样维护;设计 Skill 的本质:把品味变成约束;Skill 生态不能做成仓库列表(本质上需要强运营);内容 Skill:文章、产品和案例互相喂养;Skill 的边界会继续扩大;Skill 与 Gene:手写经验和自动进化的边界(Gene / Capsule:这里指从 Agent 反复执行中的成功路径里沉淀出的可复用经验单元,强调自动演化而不是人工手写);盗用不是靠藏,防御方式是持续分发;一个完整 Skill 生命周期;> Agent 时代最稀缺的是可复用的能力组织方式。Skill 之所以重要,是因为它第一次让人的经验、工作流和品味,有机会变成一种可以分发、调用、评价和持续迭代的商品。
-
yaojingang / yao-meta-skill: YAO = Yielding AI Outcomes. A rigorous engineering, evaluation, governance, and portability system for reusable agent skills.
-
lijigang / ljg-skills: 我的 Claude Code 自定义技能集。
-
JimLiu / Illustrated-Agent-Skills: 《图解Skill——AI提效实战指南》 官方 Repo
0)Skills 管理方案:Skills 只装在项目里,不装全局;用软链接来安装 Skills:把开源 Skills 项目下载到统一的目录 > 在自己的项目中创建软链接 > 给 Agent 建一个入口;帮我把 ~/GitHub/baoyu-skills/skills/baoyu-comic 软链接到 .agents/skills/baoyu-comic 或直接 帮我把 baoyu-skills 项目里的 baoyu-comic 这个 skill(或所有 Skills)软链接到当前项目;之前使用 SKILLS 便捷方式多 Agent 安装也采用软链接;> 20260626 逐步迁移至该 Skills 管理方案(权限原因采用 Junction 模式),Agent 根仅少量常用 Skill;注意与 Claude Code 的 Plugins - Marketplace 模式 & Codex App - Plugins 及其内置 Skill 的区分;
1)content-skills - Content generation and publishing: xhs-images, infographic, cover-image, slide-deck, comic, article-illustrator, post-to-x, post-to-wechat ; > article-illustrator 参考我写了个 Skill,让 Agent 自动给文章配图,需调用 Gemini 图像生成 Skill(如下 danger-gemini-web);comic & slide-deck 使用脚本合成 PDF/PPT;post-to-x & post-to-wechat 使用脚本通过带有 CDP 的 Chrome 绕过反自动化机制,后在 Codex 中使用 @Chrome 插件直接发送;
2)ai-generation-skills - AI-powered generation backends: image-gen, danger-gemini-web; > danger-gemini-web: 检查/创建/确认同意书 consent.json > 执行脚本 Text generation & Image generation & Vision input (reference images),运行时会打开 Chrome 并进行 Google 身份验证(另需 Proxy 设置);
3)utility-skills - Utility tools for content processing: url-to-markdown, danger-x-to-markdown, compress-image;
20260424 Skill 是天生带自杀基因的产品 - 价值创造和价值捕获被分开了 & 自杀的不只是盈利,还有数据飞轮 & 答案藏在「AI 让什么变得过剩」里 - 稀缺是关系、此刻、物理世界、以及判断与品味;
0726 plusxiaxia@plusxiaxia: 当我看到一个垂直业务智能体是一堆 Markdown 文件+一堆 Skill,我就知道,这事黄了一半。> 我们的智能体是:1 数据源;2 连接器,从数据源中提取 json string;3 本体,将 json string 反序列化为 Object;4 策略,对 Object 形成的 Graph 进行静态分析,产生 Diagnostic;5 数据绑定,将 Object 嵌入 Parametric Block 中的 Slot;6 生成,调用 Agent 补全 Generated Block,如果可追溯性非常高,我们甚至要限制 Agent 在写入特定 Generated Block 时 Scope 包含的 Block,依据哪些表达,哪些数据,做出特定判断的;7 样式绑定,将 Content Graph 中的各个 Block,根据 Target Channel,选择 Style,渲染得到 Artifact。
20251022 Claude Skills: The Functionalization of Language
1124 Claude Agent Skills:第一性原理深度解析
1229 马东锡 NLP@dongxi_nlp: Skill 是个好名字,但只是一种 trick。你可以创造自己的类似 Skill 的框架,重要的是遵循 SWE Agent 的 ACI 设计原则:1. Actions should be Simple 2. Actions should be Compact and Efficient 3. Environment Feedback should be Informative but Concise 4. Guardrails Mitigate Error Propagation > Skills 信息折叠更狠,已经几乎没有怎么启发性了。所谓自动选用,其实也是美好的愿望。最后遇到的问题会和 MCP 很像,你得明确告诉 AI,它才会过去找。一个高级工程师现查手册,肯定不如一个熟练的流水线工人效率高。> 参考 Apache ECharts vs Ant Design & AntV,后续的 SDK/API 都会自带 Skills 取代原先的手册文档,而且这种 AI 自反性会逐渐模糊 SDK/API Solidification 固化性 与 Agentic 能动性 的边界。隐约看到 SWE 3.0 的雏形,通过符号抽象升阶,让 AI 参与进来的,类似开源社区又超越了的一种涌现。> 同理,当下所有的软工开发类教材都应该以 Agent Skills 的范式编写与呈现;
20260110 当下的共识:你可能不再需要 workflow,大部分场景 skills 足矣——五步框架把 Workflow 变成可进化的 Skill
“凡是能够说清楚的,AI 早晚能够运行;凡是不能说清楚的,请保持沉默(这是人类最后的主体性)”—— 一方面是经典编程语言以及 DSL 所承载的软工终结理想的实现曙光,一方面是
1)Anthropic 专注 AI-Coding,其实深谙“代码第一性原理”,Claude Code 不仅是在快速旋转着“上下文工程(自然语言)- 软件工程(代码)”的数据/模型飞轮,在 Opus 4.5 这类推理能力到位的模型加持下,Skills 及其前后的基础设施,更是在收集迭代,从而进化为基于“大语言模型”而又超越“语言”局限的框架,这不仅仅是 MCP 阶段就呈现出的“反馈操作真实/物理世界”。我认为,是将之前“自然语言描述 vs 编程语言表达”两端间分层转化翻译职责,真正意义上模糊而频谱化,也就是真正实现软工的终极理想 ——“人类思考抽象 - 机器自主执行”的可塑性(Malleable)。回想一下之前从“写清楚需求文档”到“稳定运行代码系统”的转译过程中“文档”与“代码”本质上是两张皮的,以及针对 Vibe Coding 的 Spec-Coding 应用场景的局限性(谁不是 Specification-Driven Development?)。> 参考 Malleable Systems - Malleable Software & End-user Programming & SaaS;
成功现在取决于那些更难被自动化的因素
2.1)有人问,为何 Skills 写的都很清晰而规整(虽不如代码的精确)?因为这也是面向机器(AI)的,写不清晰跑不出效果(当然 AI 的自反性会方便澄清与迭代),而一旦 AI 掌握的这个技能,自主组合后的收益是巨大的,尤其是相较之前需要人工编程转译所投入的成本而言,构建与迭代成本是极大下降的。这是一个人机协作的可塑界面,对一个面向 AI 的写作者/编程者而言,他当然可以继续指望后面其他人来澄清
20251019 部署调试 Agent Skills,网页版/桌面版 Claude 需 Pro 账号,故直接调试等效开源的 anthropics / skills(安装 Marketplace 时 Git 至用户目录):
0)marketplace.json 配置了两个 Plugin:document-skills 对应 document-skills;example-skills 对应其他子目录的技能;> 1220 新版本结构调整,更新 marketplaces & plugins:skills 目录下 pdf、pptx & docx & xlsx、frontend-design;
1)调试 document-skills,创建一个两页的介绍 Claude Code Skill 的 PPT:在目录 document-skills-demo 中,构建 slide1.html & slide2.html > 构建 create-gradient.js 生成背景 gradient-bg.png > 构建 create-presentation.js,运行时需拷贝脚本 html2pptx.js > 就近安装 Node 依赖库(需关注运行环境及沙箱,见下文)> 构建 PPT 过程中,反复修改 .html 文件(GLM-4.6 能力问题)> 构建成功,生成内容 HTML 以及 Layout 等进而 NodeJS 脚本构建 PPT 的技能固化,而由 CC 具体执行并调整;
2.1)调试 skill-creator,引导我创建一个 Skill,实现对 csv 文件中的数据进行各种分类的机器学习分析,采用 scikit-learn 库,参考 Hands-On Machine Learning with Scikit-Learn and TensorFlow - 机器学习实战 & Machine Learning Bookcamp - Python 机器学习项目实战:载入 skill-creater 技能,分析并设计核心功能模块任务,在目录 skill-creater-demo/csv-classification-ml 中开始创建技能:SKILL.md & requirements.txt & README.md 等;scripts/ 各步骤 Python 脚本 & assets/ 数据资料 & references/ 算法文档 > test_skill.py & simple_test.py 仅检查技能目录文件结构 > 可直接执行如 python scripts/csv_classifier.py assets/example_datasets/iris_classification.csv species;
2.2) 在 skill-creater-demo 目录中配置 marketplace.json(参考上文)为 ml-agent-skills,而 csv-classification-ml 等目录即为 data-science-skills 技能组之一;CC 中安装 Marketplace & Plugins 后调试该技能:IDE 插件中采用 /csv-classification-ml 而没有 /plugins(显式 Plugins);
-
使用样例数据 iris_classification.csv 进行机器学习分类;过程脚本放在目录 ds-skills-test 中(注意 scripts & models 等子目录结构),运行采用 Python 虚拟环境 C:\ProgramData\anaconda3\envs\homl3:载入 csv-classification-ml 技能,分析并设计核心功能模块任务 > 构建 scripts & results & plots & models & data 目录结构,以及 iris_classification.py 等脚本(为新生成的代码)> 调整 Python 虚拟环境,并检查依赖库 > 运行脚本成功,并写入结果; -
使用样例数据 titanic_classification.csv 进行机器学习分类,请尽量调用原 Skill 中已有的函数;……:先行检查 homl3 中的依赖库是否已满足 > 载入 csv-classification-ml 技能,分析并设计核心功能模块任务 > 将样例数据以及技能脚本拷贝至工作目录 > 构建 titanic_analysis.py 脚本(调用技能脚本)> 安装虚拟环境依赖库,确认都已满足 > 运行功能脚本,数据预处理 - 特征工程 - 模型训练 - 模型评估 - 结果可视化,不断调整脚本(含少量拷贝的技能脚本,涉及文件路径等),成功后结果写入;
1221 部署使用 thedotmack / claude-mem & obra / superpowers & muratcankoylan / Agent-Skills-for-Context-Engineering:注意 plugins 目录中 marketplaces 更新源码而 cache 中动态载入;
1)/plugins 安装 claude-mem 后重启 Claude Code 报错:SessionStart:startup says: Plugin hook error:📝 Claude-Mem Context Loaded ℹ️ Note: This appears as stderr but is informational only > Web 端 http://localhost:37777/ 正常;过程内容参考 ~/.claude-mem;> 后运行时弹 Node 命令框且阻塞(后确认为正常),故暂关闭;> 0113 更新后正常;
2)部署调试 Agent-Skills-for-Context-Engineering & obra / superpowers(202605 Codex 调试,详见 SuperClaude Framework & SuperPowers);
20260121 调试 JimLiu / baoyu-skills :针对 danger-gemini-web,使用头部模型可以很好遵循 Skill 意图执行脚本,Google 用户认证需系统 PIN 码,并确保 Gemini Pro 可连接;> 后 Skills 在 Claude Code 中安装,而在 OpenCode - Zen / OMO(LazyCodex) & Crush 中调试(充分利用各类模型配额);

