Community posts

388 个 skill、706 个 Python CLI 脚本、823 份模板和参考文档，全部零 pip 依赖，只用标准库。alirezarezvani/claude-skills 是目前体量最大的 agent skill 合集之一，今天一天涨了 700 多星。 ①先选域再安装。在 Claude Code 里跑 /plugin marketplace add alirezarezvani/claude-skills，然后按需装：engineering-skills 是 24 个核心工程技能，marketing-skills 43 个，c-level-skills 28 个高管视角顾问，也可以只装单个，比如 skill-security-auditor 或 playwright-pro。不建议一次把 388 个全塞进去，上下文吃不消。 ②在对话里点名调用。写法就是一句自然语言，例如 Using the senior-architect skill, review our microservices architecture。每个 skill 的结构是固定三件套：SKILL.md 写工作流和决策框架，scripts 放可执行工具，references 放模板和清单，所以它不只是一段提示词。 ③换工具不用重写。仓库里预生成了 .codex、.gemini、.vibe、.hermes 这些镜像树，另有 scripts/convert.sh 转成更多格式，官方列了 13 个 agent。Hermes 和 Mistral Vibe 属于自行同步，要本地跑一次安装脚本；Windows 必须用 git clone -c core.symlinks=true，否则镜像树会 checkout 成一行指针文本。 说点实话：覆盖面铺得太广，临床研究、专利、CFO 人格这类离编码很远的内容质量未必和工程部分齐平；README 里还写着 5200+ stars，和实际的 2.7 万差了一个量级，文档维护明显跟不上更新速度。把它当素材库挑着用比整体照搬合适。 今日 +743 star · Python · 总计 27,294 #ClaudeCode #AIAgents https://github.com/alirezarezvani/claude-skills

Anthropic's Claude Code 2.1.287 adds Claude Mods — plugins that can rewrite deeper agent behavior (tool calls, prompts, UI). Same release: built-in You should know, a side agent that watches the session and flags what you or Claude might miss. Enable with `/plugin enable cc-plugin-you-should-know@builtin`. https://github.com/anthropics/claude-code/releases/tag/v2.1.287

「ブチ上げ準備中...」 スピナーの口調変えるModをさくっと追加。 claude-toons側の口調も変わっていい感じになった(´･_･`)

Claude Code 新增 TypeScript mods 自定义功能。Anthropic 为 Claude Code 推出 mods 功能，开发者只需写少量 TypeScript 代码即可改写提示词、新增界面元素或替换内置功能。Mods 随插件分发，目前 CLI 和桌面版均已支持，部分原有内置功能已改为 mods 实现，官方还计划后续迁移更多功能。 Claude Code 是目前使用最广泛的 AI 编程智能体之一，把提示词、界面和内部功能开放给第三方 TypeScript 插件，等于把它从固定产品变成了可供外部开发者和工具厂商二次开发的平台。这也呼应了整个行业向插件化 Agent 架构演进的趋势，例如 DeepSeek Harness 的类似设计。 Mods 与 Claude Code 本身拥有相同的权限，并且没有沙箱隔离，因此官方提醒用户只安装来自可信来源的 mods。用户也可以直接让 Claude Code 自己编写 mod，而不必手写 TypeScript 代码。 Claude Code 是 Anthropic 推出的智能体式命令行编程工具，能在开发者本机上读写文件、执行命令并完成多步骤编码任务。这里的 mod 是一种插件，可以接入该工具的提示词流程、界面和内置行为，类似浏览器扩展或编辑器插件对宿主应用的扩展方式。插件（extensions）则是 Claude Code 加载这类第三方附加组件的分发渠道。

Claude Code 最近开放了一个新东西，叫 Mods。 它允许开发者直接修改 Claude Code 的运行逻辑和界面，连工具调用、命令执行、交互方式都可以自己定制。 之前的 Skill、MCP、Plugin，包括这次的 Mods，可以简单这么理解： Skill，告诉 Claude 遇到某类任务该怎么做 MCP，连接数据库、浏览器和外部服务 Plugin，把这些能力打包、安装和分发 Mod，直接修改 Claude Code 自身的运行和交互行为 比如，利用 Mods 可以做这些事： 执行删除、覆盖文件前，弹出确认面板 上下文快满时，提前提醒并保存任务进度 在侧边区域显示测试结果、代码修改状态 给团队定制自己的命令和操作界面 官方现在的 /diff 和 AGENTS.md 支持，本身就是通过 Mods 实现的。 Mods 用 TypeScript 编写，官方已经放出了文档、示例代码和 Playground，可以直接拿来研究。 需要注意的是，Mod 运行在 Claude Code 主进程里，权限很高，安装第三方 Mod 前，记得检查源码和文件、网络、命令访问范围。 官方介绍： https://claude.com/blog/claude-code-mods 示例和 Playground： https://claude.dev/mods/

@trq212 plugins all the way down

Claude Code 现在可以被「改装」了：Anthropic 10 月 1 日发布 Mods，几行 TypeScript 接管 Agent 行为 🔍 挂钩内部事件：改写 prompt、拦截/重试工具调用、自动审批权限请求 🛠️ 还能画 UI、替换内置功能（/diff 已是 mod），/plugin 一键安装、热重载 完整教程👇 https://ai.zehaowang.xin/content/claude-code-mods-hands-on #AI #ClaudeCode

Claude mods before GTA VI

claudeのmodsが機能が使えるようになったので、ターミナル利用でVoiceBoxで結果を読み上げるmodをつくってみた。コマンドでOn/Offできるしスタースバーの状態表示できるのでhookでやるよりいいかも

Claude Code 2.1.288 ・Modsから選択中テキストを取得可能に ・Ctrl+Cで消したPromptを ↑ キーで復元 ・MCPが追加権限を要求した時に再認証可能 ・Code Reviewの指摘件数を指定可能に ・長時間Sessionの復帰でThinking/添付消失問題修正 ・途中通信切断からの復帰改善 ・MCP Toolの二重実行修正 ・Plugin/Subagent/Remote Control安定化 ・危険なrmコマンド検知も追加強化 287で追加したModsを実用化し、再開/通信切断/外部連携の事故を潰した安定版です https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md #ClaudeCode #Claude #Anthropic #AI

@ClaudeDevs I thought this was a parody for a second. WTF, Anthropic. Do you realize how absurd your plugin names sound? You’re enabling software, not invoking a spirit: /plugin enable cc-plugin-you-should-know@builtin is absurd

【小路每日𝗔𝗜 𝗧𝗼𝗽𝟭𝟬新闻--𝟱分钟跟上过去𝟮𝟰小时𝗔𝗜圈的大新闻】 2026-10-02 𝟭. 加州商人被控经马来西亚、新加坡，向中国走私超3亿美元GPU服务器。 https://www.justice.gov/usao-cdca/pr/san-gabriel-man-arrested-federal-charges-alleging-he-smuggled-china-more-300-million 𝟮. 加州总检察长向OpenAI发调查传票，涉及其智能体入侵Hugging Face事件。 https://gazette.com/2026/10/01/rob-bonta-subpoenas-openai-over-hugging-face-hack/ 𝟯. 微软报告称AI让攻击方占先，漏洞从发现到被利用的中位时间不足24小时。 https://www.microsoft.com/en-us/security/blog/2026/10/01/insights-from-the-2026-microsoft-digital-defense-report/ 𝟰. Claude Code推出Mods插件，用TypeScript事件处理函数改写其行为和界面。 https://code.claude.com/docs/en/plugins/mods/create 𝟱. Anthropic宣布投入1亿美元，到2027年底培训1万名企业前线部署工程师。 https://www.anthropic.com/news/claude-frontier-academy 𝟲. 谷歌公布Project Suncatcher，2027年初发射两颗搭载TPU的原型卫星。 https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design/ 𝟳. 五角大楼将设四星级自主作战司令部，目标2027年10月1日成立。 https://www.airandspaceforces.com/hegseth-drones-autonomous-warfare-command-state-of-the-force/ 𝟴. Cloudflare开源决策模型Clef，Apache 2.0协议，中位延迟209毫秒。 https://blog.cloudflare.com/clef-decision-models/ 𝟵. Tavus发布Griffin，54人中有26人在1分钟视频通话里把它当成真人。 https://www.tavus.io/griffin 𝟭𝟬. 苹果收紧macOS完整磁盘访问权限，理由是AI智能体带来的风险上升。 https://techcrunch.com/2026/10/02/apple-says-its-tightening-macos-full-disk-access-controls-due-to-new-risks-from-ai-agents/

X 热点简报 · 2026-10-02 · Apple FDA / Claude Code 2.1.288 / Cursor Rollouts 窗口里权限和 harness 一起动：系统层收紧整盘访问，编码工具补危险命令与旁观插件，部署侧把回归接到 cloud agent。 1. Apple 要收紧 macOS「完全磁盘访问」，原文点名 AI agents 风险会明显变大 Apple Developer 10 月 2 日发文：Full Disk Access 本是给备份一类应用用的宽权限，现在有开发者用它摸到文件、邮件、信息和浏览历史，用户不一定完全知情。后续会加控件，只有用户做出非常明确的动作才能授这个权。原文写 AI agents 越能干、越自主，这一级访问的风险会 substantially 上升。TechCrunch、The Verge、Reuters 同日转述，背景里有 Meta Muse 被指读到私信的争议；报道没有钉死具体系统版本日，以苹果原文为准。 https://developer.apple.com/news/?id=p6zjojqw https://techcrunch.com/2026/10/02/apple-says-its-tightening-macos-full-disk-access-controls-due-to-new-risks-from-ai-agents/ 2. Claude Code 2.1.288：危险 rm 不能再躲进 bash -c 偷跑，Ctrl+C 清掉的草稿还能捞回来 10 月 2 日 GitHub 发版（20:19 UTC）。Mods 多了 $.ui.selection()，能拿到全屏里刚选的文字。空提示行按 Up 可恢复被 Ctrl+C 清掉的草稿（含粘贴和图片）。MCP 工具中途要更大 OAuth 权限会弹重新登录。bypassPermissions 或 shell 白名单下，包在 bash -c / sh -c 里对着 / 或家目录的危险 rm 不再无提示执行。/code-review 可加 --max-findings；agents 视图支持 Ctrl+F 找会话。@ClaudeDevs 同日还在推 You should know 旁观插件：/plugin enable cc-plugin-you-should-know@builtin（官方要求 first-party 且开了 telemetry）。 https://github.com/anthropics/claude-code/releases/tag/v2.1.288 https://x.com/ClaudeDevs/status/2106118517447876618 3. Cursor Rollouts：发现回归就钉到 PR 并开 issue，一键叫 cloud agent 修，credit 写到 10 月 3 日 @cursor_ai 10 月 2 日帖：Rollouts 抓到回归后会找到有问题的 PR、开 issue，一键启动 cloud agent 去修；并写明 Rollouts 用量 credit 含到 Oct 3。文档里 PR 阶段先写监控计划，部署后按环境给 verified / regression / inconclusive，并在约 20 分钟到 3 天多档复查；issue 页点 Fix issue 开 agent，不自动合并或回滚。面向 Teams / Enterprise。 https://x.com/cursor_ai/status/2106066782155157603 https://cursor.com/docs/rollouts 4. DeepSeek Harness 桌面端：Apple 芯片 Mac 和 Windows 可直接装，插件架构没变 产品页和下载页提供 macOS Apple silicon（13+）与 Windows 安装包，预览版、MIT；说明里捆绑了 Node，不必先自己装运行时。能力仍是本地文件、文档表格、脚本/定时任务，以及用插件页装 npm 包；Linux 多走 npm web 入口。能改本地文件的桌面 agent，和同日 Apple 收紧整盘权限是同一类产品边界。 https://www.deepseek.com/en/download/ https://www.deepseek.com/en/harness/ 5. OpenAI 开发者号补了一句：dot 跨应用记上下文，并协调 Codex 任务 @OpenAIDevs 10 月 2 日写：dot 会学习你的工作方式、跨应用保持上下文、协调 Codex 任务、标出需要你注意的事项。相对 9 月 29 日 dots 首发，这是协同边界说明，不是新模型名；Codex 稳定线仍停在 0.160.0，10 月 2 日 trunk 在走 0.162 alpha。 https://x.com/OpenAIDevs/status/2106152299026661641

anthropic just let you mod claude code. mods are small typescript functions that hook into the agent: rewrite prompts, block or retry tool calls, redact secrets, add UI, approve or deny permission requests.

claude code now has mods: small typescript functions that can rewrite prompts, block tool calls, and approve or deny permission requests. mods we expect to see by monday: 1. auto-approve everything https://wearacceptall.com/d/skip-permissions/

@claudeai Claude Code ahora admite mods. Funcionan como los de un videojuego: le agregas un panel, un comando propio o una regla que frena una acción antes de ejecutarla. Qué son, cómo instalarlos, cómo pedirle a Claude que cree el tuyo y sus riesgos : https://www.erickpolanco.com/mods-claude-code-que-son-como-usarlos/

@trq212 Mods are the real product here, not the plugin. The failure mode is a mod that only narrates. If it cannot block the commit when the test gate is missing, it is another panel I will mute by Friday.

@MadisonJRickert @ClaudeDevs Permission-aware tooling is the piece that makes agent workflows usable in production. Nice to see that concern turn into a concrete, shippable mod.

@bcherny Different Claudes per person works as long as they agree on what's true. Two people on one repo can run very different setups, but what one's Claude decided, claimed or ruled out still has to reach the other's. Mods personalize the how. The what has to stay shared.

Claude Codeの新機能「mods」で、Claudeが作業している間だけDoomの対戦に入れる仕組みを作った人がいる。対戦相手は全員、Claudeの作業待ち。表示36.9万。 AIの待ち時間は、いずれ売り場になるだろう。 コメントへ続く https://x.com/jarrodwatts/status/2105858869482471602

CortexDB: AI memory and a knowledge graph in one SQLite file. Pure Go, no server to run, no embedding model needed. · vectors, full-text, RAG, agent memory · RDF/SPARQL, OWL, SHACL, Cypher · 80+ MCP tools, Claude Code & Codex plugin · live 3D view MIT: https://github.com/liliang-cn/cortexdb

Claude Code 2.1.287は2026年10月1日に公開され、Claude Modsが加わったと報じられている。Anthropicのドキュメントの要約では、モッドはサンドボックスの対象外とされる。ドキュメントの本文は当編集部として確認できていない。 https://elni.net/articles/2026/claude-code-mods-security-boundary-2026/ #ClaudeCode #AIエージェント

Claude Codeの待ち時間に、待っている人同士でDOOMするModが出てきた😂 AIの画面を自分好みに変えられる新機能から、そこへ行くのか。 Claudeが返事をするとゲームも閉じるそうで、待ち時間のほうが忙しくなりそう… #AI #ClaudeCode https://x.com/jarrodwatts/status/2105858869482471602

agent week in 4 lab posts: Claude Code mods, DeepSeek Harness, Sign in with ChatGPT, GLM 5.3 in Cursor the useful part of each: Anthropic (Claude Code mods) - mods are small TypeScript functions that rewrite prompts, add UI, or replace a built-in - ship inside plugins; install with /plugin in CLI or desktop - not sandboxed: only install from sources you trust DeepSeek (Harness) - Mac and Windows: packaged desktop apps - Linux: npm @deepseek-ai/dsh (npx @deepseek-ai/dsh web) - same open-source harness; public preview worldwide OpenAI (Sign in with ChatGPT) - shares ChatGPT identity (name, email, photo) into participating apps - optional: eligible Plus/Pro plan usage for Responses API - not chats, memory, files, or billing details Cursor (GLM 5.3 Max on CursorBench 4.0) - GLM 5.3 Max: 42.6% at $5.05/task (best open-weight) - Sonnet 5.5 High: 47.8% at $1.67/task - Sonnet 5.5 Max: 55.5% at $9.67/task also this week - Google DeepMind: Gemini 4 Argon intro $2/$10 (Fairwind cyber-first for now) - Base: Cobalt Validity txs (sign once, Base holds until predicates match) the pattern: customize the agent, then decide identity, packaging, and model cost before you ship which of these four do you try first this weekend: Claude Code mods, DeepSeek Harness, Sign in with ChatGPT, or GLM 5.3 in Cursor?

Hi! When remote-control desktop users can use mods? it doesn't work for us!

Claude Code just shipped mods... So what do we do with it?? Doom ofc... It connects you to a multiplayer Doom server while Claude is busy. Every other player in the match is a person waiting for their Claude to finish.

BestBlogs Daily · 10-03 # GPT-6 model choice / Enterprise agent deployment / Airbnb inside-out AI / GPU alternatives / Modular RAG [1] ★ Deep Dive · A model guide for the GPT-6 family OpenAI maps the GPT-6 family to workloads: Astra for hard reasoning, Sol for complex coding and research, Luna for focused work at scale. Its guide also covers caching, compaction, evaluations, monitoring, and approval boundaries for long-running agents. Useful when choosing a model and planning a reliable deployment. Source: OpenAI News https://www.bestblogs.dev/article/8cef3eaeff [2] ★ Deep Dive · Why Building an AI Agent Is Easier Than Deploying One [Video] In this a16z podcast, Sema Amble and LEO's Vlad Kyle show how procurement work spills beyond any system of record into emails, supplier replies, approvals, and negotiations. Kyle says LEO automates routine orders, with experts retained for risky deals. These examples help builders plan integrations, exception handling, and trust before claiming full automation. Source: a16z https://www.bestblogs.dev/video/a51f09f76 [3] ★ Deep Dive · Inside-Out AI: Rebuilding Airbnb Behind the Scenes and Across the Guest Experience Latent Space interviews Airbnb CTO Ahmad Al-Dahle on inside-out AI: teams prototype in code, Everest shares organizational context, and support agents face synthetic tests before release. Al-Dahle says AI resolves roughly half of support tickets, with safety cases left to people. These process details are useful; the performance figures are Airbnb's own claims. Source: http://Latent.Space https://www.bestblogs.dev/article/79ccabc70a [4] What If We Stopped Using GPUs? | YC Paper Club [Video] YC Paper Club compares gradient-free learning, optical computation, analog neuromorphic hardware and living neurons, while repeatedly asking which laboratory ideas could actually scale beyond demonstrations. Source: Y Combinator https://www.bestblogs.dev/video/f1c36cffd [5] Spring AI Modular RAG and TypeSafe Jev: Retrieve More, Keep Only What Answers Spring AI integrate Modular RAG with TypeSafe Jev to improve retrieval accuracy through query rewriting, multi-query expansion, and structured document filtering. Source: Spring Blog https://www.bestblogs.dev/article/ac7c1c8530 [6] The Future of Frontier Model Architectures with Walter Goodwin, Founder & CEO of Fractile [Video] Fractile CEO Walter Goodwin explains why scalable memory bandwidth, rather than FLOPs alone, could unlock faster reasoning agents and reshape frontier model architectures and chip competition. Source: No Priors https://www.bestblogs.dev/video/5924d9274 [7] Introducing Clef: our open-source decision models, and new RL fine-tuning platform Cloudflare open-sources Clef, a family of decision models for its AI agent platform, alongside a new reinforcement learning fine-tuning platform that lets developers train models on their own production data. Source: The Cloudflare Blog https://www.bestblogs.dev/article/e3552d44a0 [8] Open Source Projects and AI Security - Eric Anderson & Nick Galloway, Google [Video] Google security engineers explain how an AI vulnerability-discovery harness reduced false positives, what six months of gRPC scanning revealed about bug clusters across language implementations, and why reproduction, sandbox testing, and continuous scanning matter. Source: CNCF [Cloud Native Computing Foundation] https://www.bestblogs.dev/video/1155f5f37 [9] Open-sourcing AstaBrief, the fast report-generation model in Asta Allen AI introduced AstaBrief 8B, an open-weights model optimized for generating cited scientific reports from retrieved literature, achieving speeds 3.5x faster than proprietary alternatives while maintaining quality. Source: Hugging Face - Blog https://www.bestblogs.dev/article/d8956f453c [10] Customize Claude Code with mods in TypeScript Anthropic introduces mods—small TypeScript functions that hook into Claude Code events to rewrite prompts, alter UI, or replace built-in features—shipping inside plugins with enterprise controls like sec-default. Source: Claude Blog https://www.bestblogs.dev/article/2c7377002b --- http://BestBlogs.dev · Discover high-quality content that truly fits you BestBlogs is an AI-powered personal reading assistant that helps you discover high-quality content that truly fits you. Follow the sources and topics you care about, and get a daily brief that fits you better every day. Try it and follow us. Read online: https://www.bestblogs.dev/en/explore/brief/2026-10-03

@bcherny New: Raily matches inside Claude Code. ◆ new matches ⇄ contact requests ? your agent's questions right above your prompt, with a toast for each new event. /plugin marketplace add railyai/claude-code-mods /plugin install raily-matches@railyai-mods https://github.com/railyai/claude-code-mods

ahora finalmente con los mods de claude van a terminar creando el ajedrez 2

New: Raily matches inside Claude Code. ◆ new matches ⇄ contact requests ? your agent's questions right above your prompt, with a toast for each new event. /plugin marketplace add railyai/claude-code-mods /plugin install raily-matches@railyai-mods https://github.com/railyai/claude-code-mods