Community posts

pstack es el stack de skills que usa lauren tan en Cursor para que el agente trabaje mejor. ahora alguien lo ha portado a Claude Code, Codex, Pi, OpenCode y Gemini CLI. la idea: le dices a poteto-mode qué quieres conseguir y él elige el workflow que toca. por ejemplo, con un bug: 1) reproduce el fallo 2) investiga el porqué 3) delega el fix 4) vuelve a ejecutar el caso que fallaba y te devuelve el fix con la prueba del antes y el después. si el cambio cruza más de una función, mete a un agente architect antes de implementar. también trae playbooks para planificar, refactorizar, rendimiento, prototipos, PRs y proyectos largos. en Claude Code se instala con dos comandos: /plugin marketplace add michael-denyer/pstack-claude /plugin install pstack@pstack-claude

有人把 Claude Code 的插件机制（Mods）玩出了新花样：直接在终端输入框上方常驻养了一只像素宝可梦。 这不仅是个怀旧像素彩蛋，而是把 Agent 的底层状态和生命周期做成了电子宠物玩具： - 当 Claude 在后台跑代码或检索时，宝可梦会在终端上方实时走动； - 它的思考气泡会实时同步显示 Claude 当前正在调用哪一个底层工具（Tool）； - 只要你派出一个 Subagent 子代理，它就会在终端里丢出一颗精灵球，任务搞定精灵球自动弹出； - 你每跟 Claude 进行一轮交互，它就能累积经验值（XP），按照红蓝宝石原版的等级规则真实进化。

claude mods ありがとう。 大変使いやすいです

Claude Code 2.1.289 closed a real sandbox hole. Deny rules on nested shell commands could be overridden by a mod approval. An env prefix like FOO=bar also skipped bash deny rules under auto-allow. If you sandbox, update. #ClaudeCode

いいな / 【Claude Mods】4つの自作 Mod が神連携(Treg/Jev/GBrain/RTK/Recall) https://youtu.be/cIfmFBM_tbY?si=KoSyI6-bYhEsGWeR @YouTubeより

Claude Code mods give u your own creativity within claude. You can change how Claude behaves, add your own UI, block commands, rewrite prompts, or build completely new features. So naturally, my first thought was: what ridiculous mod should I build? https://surl.li/lbgysa

Claude Codeを書き換えるMods登場🔧 ✔ TypeScriptで挙動も画面も変更 ✔ /pluginで導入できる ✔ ただしサンドボックスなし 詳しくはこちら👇 https://ai.seater009.com/posts/2026-10-05-2026101claude-codemodstypescriptmcp/ #AI #生成AI #AIニュース #ClaudeCode

Claude Code just got MODs Small TypeScript functions that hook into Claude Code itself They can: block or rewrite a tool call rewrite your prompt before it hits the model redact secrets from output replace built in features like /diff No sandbox though. A mod runs with full access to your machine Install one from a stranger and you just gave them your computer Full control, real risk. That is the tradeoff

Claude Code doesn't show your token usage on screen. You have to go and look for it somewhere else. So I asked Claude to fix it with a mod. Now daily and weekly token consumption sit above my prompt, in real time. One color per category. What else would you put on the bar?

#ClaudeCode Mods let devs rewrite the tool from inside. ⚠️ No sandbox + your permissions = one bad Mod owns your machine. 🧩 JS/TS event hooks make it endlessly hackable. 🔐 Anthropic itself says trusted sources only. Listen. https://cryptogames.gg/claude-codes-new-mods-let-developers-rewrite-the-ai-coding-tool-from-inside/ #Anthropic

The usage dashboard I wrote for Oh My Pi now runs in Claude Code, built on its new mods: token chart, local history and 5h/7d limits in a native Pane. Live accounting is not verified yet. https://mspiechowicz.com/blog/harness-usage-dashboard-claude-code

@lydiahallie Built a secret mod it's good I saw one more thing idle session compaction on one of my Claude subs but it's not enable overall Is there anyway you can roll it to all or add a flag so users can enable that?

Vasat, yani ortalama kodlama araçlarına daha iyi orkestrasyon sağlayacağı ümidiyle mod, skill, plugin, talimat ve MCP eklemekten bunalanlara günaydın. 😄 Claude Code, Codex, Pi, OpenCode, Antigravity, CommandCode… Tercihiniz hangisiyse artık. Birini kuruyor, eksiklerini fark ediyor, sonra o eksikleri başka bir şeyle kapatmaya çalışıyorsunuz. Biraz hafıza, biraz subagent, biraz review, biraz da “bu sefer talimatları unutmasın” diye yeni bir .md dosyası. Derken proje geliştirmekle kendi kodlama aracınıza bakıcılık yapmak birbirine karışıyor. WrongStack tarafında ise hafıza, görev yönetimi, model yönlendirme, araç kullanımı, ajanlar arası iletişim ve doğrulama aynı motorun içinde birlikte çalışıyor. Kendi çekirdeği, provider bağlantıları, tool executor’ı ve çok ajanlı çalışma sistemi var. Baştan sona bağımsız bir geliştirme ortamından bahsediyorum. Dahili plugin ekosisteminde deterministik çalışan yetenekler de var, gerektiğinde LLM kullanan yetenekler de. Karar tarafında Brain’in kuralları, isteğe bağlı Jev değerlendirmeleri ve daha yüksek riskli konularda çok modelli Council devreye girebiliyor. Üstelik pluginlerin modelini de ayrı yönlendirebiliyorsunuz; her küçük iş için liderin kullandığı modeli çalıştırmak zorunda değilsiniz. SAGE memory sisteminde bir hafıza kaydı dosyaya, dizine, komuta, teste, pakete, commit’e veya doğrudan bir class’ın içindeki metoda kadar bağlanabiliyor. “Bu projede bir ara böyle bir şey yapmıştık” seviyesinde hatırlamadan bahsetmiyorum. Kararın hangi kodla ilişkili olduğu belli, dayandığı kaynak değiştiğinde geçerliliği yeniden kontrol ediliyor. Memory Companion da kullanılan hafızanın mevcut kodla uyumunu denetliyor. Kanban tarafında görevlerin bağımlılıkları, sorumluları, aşamaları ve kabul kriterleri var. Ajan “bitirdim” dediğinde kart otomatik olarak Done olmuyor; Review aşamasına geliyor, tamamlanma koşulları ve kanıtları doğrulanıyor. Çünkü kod yazmış olmakla işi tamamlamış olmak aynı şey değil. Aynı oturumda lideri bir provider/model ile, reviewer’ı başka biriyle, bug hunter’ı başka biriyle çalıştırabiliyorsunuz. Rol ve aşama bazında routing, ayrı fallback profilleri ve uygun provider hatalarında otomatik geçiş var. Bir modelin kapasite sorunu bütün çalışma düzenini tek başına rehin almasın diye. Codebase indexleme sayesinde ajan dosyaların içinde deli dana gibi dolaşmak yerine sembolleri, çağrı ilişkilerini ve ilgili kodu arayabiliyor. Uzman ajanlar bulgularını proje çapındaki mailbox üzerinden paylaşabiliyor. Ayrı oturumlar ve worktree’ler arasında da iletişim var; herkesin birbirinden habersiz aynı sorunu yeniden keşfetmesine gerek kalmıyor. ToolFlow ile birbirine bağlı araç çağrıları bir akışta işlenebiliyor; ara çıktılar gerektiği yerde süzülüp modele karar için gereken sonuç aktarılıyor. Chimera değişiklikleri inceleyip dosya ve satır referanslarıyla bulgu üretiyor. Session Story ise hangi ajanın hangi araçla ne yaptığını, hangi dosyaya dokunduğunu, hangi hafızayı kullandığını görünür hale getiriyor. Ne olduğunu tahmin etmek yerine kaydını inceleyebiliyorsunuz. Bir de uzman ajanların proje özelinde öğrendiği çalışma biçimleri var. Bir reviewer’ın veya bug hunter’ın bu repoda işe yarayan deneyimi sonraki uygun görevde tekrar kullanılabiliyor. Her yeni session’da aynı şeyleri sıfırdan anlatmak zorunda kalmayın diye. Sonra bir yerde “model aptallaştı”, “eskiden daha hızlıydı”, “bu sefer projeyi bozdu” cümlelerini görüyorum. Peki sorun gerçekten modelde mi? Modele hangi bağlam verildi, hangi araçlar sunuldu, eski bilgi nasıl ayıklandı, görev nasıl bölündü, sonuç nasıl doğrulandı? Bunlara bakmadan bütün faturayı modele kesmek ne kadar anlamlı? “Al kardeş, proje burada. Kafana göre takıl; bozulursa ben sana fixle diye prompt atarım” düzeninde çalışıp sonra bunu güçlü bir mühendislik sistemi sanmaya fazla alıştığınızı düşünüyorum. Benim vasat dediğim tam olarak bu. Birkaç iyi task çıktısından sonra aracın proje büyüdükçe de aynı kaliteyi koruyacağını varsaymak. Eksik kalan her şeyi yeni bir talimat dosyasıyla kapatmaya çalışmak. Sonra o talimatları da sizin takip etmeniz, hatırlatmanız ve denetlemeniz. WrongStack her gün bu zincirin başka bir halkasını geliştiriyor. Ben aradaki farkı hafızanın, araçların, ajanların ve doğrulamanın proje boyunca birlikte çalışmasında görüyorum. Kullandığınız araç beklentinizi karşılıyorsa güzel. Ama beklentinizi aracın sınırlarına göre küçültmüş olabilir misiniz, onu da bir düşünün.

Claude Code can now be modded from the inside. Powerful, and it comes with one rule worth reading before you install anything 🧵 #ClaudeCode

I want my model recommender to actually pick the model, not just suggest one. Hooks can't do that, mods might. Routing one Claude Code request to another model (not tried yet): https://losoft.dev/notes/route-one-claude-code-request/

i wanted to change how claude code looked long before i understood what it was doing. now you can. anthropic shipped mods: change how it behaves, customize the ui, swap in your own features. a few lines of typescript, and the mod ships inside a plugin, so you install it with /plugin. the uncomfortable part is that the look was never what cost me anything. what cost me was underneath it. i spawned claude -p from /tmp once and it broke my mcp server registration without throwing a single error. a nicer pane would not have told me that. i'll still write one. a status line that says which of my agents are alive, because today i find that out by opening sqlite and reading the table myself.

Claude Code can now be modded. A few lines of TypeScript can block a command before it runs, rewrite a prompt or redraw the UI. Mods are not sandboxed. Same access to your machine as Claude Code. Install only from trusted sources.

Saving Tokens in Claude Code Cache Control for Claude Code. This is probably one of those mods that helps save tokens the most and keeps sessions alive noticeably longer. It adds a separate bar for Claude's 5-minute cache and sends notifications to help you keep the cache warm. Install command: npx claude-code-templates@latest --mod observability/prompt-cache-control

Claude Codeのmod、キャッシュ5分の残りがバーで見えるのが新しい👀 長く席を外す前にバーを確認し、切れそうなら区切りのいいところまで先に進める。今日試せる一歩はこれ 長時間回す人向け https://x.com/dani_avila7/status/2106455605925822967

Anthropic の Claude Code の mod 入門記事です。 mod は JS/TS で hook を定義する Plugin で、通常の hook と違いセッションに常駐し、状態保持や UI 描画もできます。 hook の動作の型 (Observe / Rewrite / Answer) や chain、サンプル実装の解説がわかりやすいです。 https://claude.dev/blog/getting-started-with-claude-code-mods/

loving the new claude code mods feature @ClaudeDevs @orca_build

Claude Code v2.1.289、更新直後にMODが読み込まれない不具合を修正 Claude Code 公式CHANGELOG の v2.1.289 で、不具合の修正が1件公開されました。 内容は「アップグレード後の最初のセッションで、インストール済みの mods（拡張機能）が読み込まれない」問題の修正です。 バージョン更新の直後だけ挙動が変わる、という種類の不具合でした。 私も7月16日の朝、定期実行6本中5本が止まったとき、AIが「許可設定の構造的な欠陥だ」と診断してきたのを止めて、「昨日まで動いていたものがなぜ止まったのか、が起点では」と指摘しました。調べたら前夜にClaude Codeが自動更新されていたのが原因で、更新直後は『設定ではなくバージョン差分を先に疑う』のが近道です。 出典: Claude Code 公式CHANGELOG

Claude Code のデスクトップ版に Mod で手を入れてる。 入力欄の上に 5時間/7日の使用量メーターと effort の切り替え、 考えてる間はドット絵の俺が7つの場面を走り抜ける劇場。 実際の画面録画です👇

Anthropic just shipped Mods, TypeScript hooks that can rewrite Claude Code mid-flight. I want every agent runtime to grow a first-in-chain deny layer before marketplace plugins get a vote. Config knobs were never going to hold a team that installs hot-reload middleware on the same machine as the model. Claude Code now turns plugins into event hooks that rewrite prompts, mutate tool calls, gate permissions, redact output, and draw custom UI, and Team/Enterprise get a sec-default mod.

これめっちゃ便利！ 次に取れるアクションを提案してくれて、選択すると指示を補完してくれる インストールはこちら👇 claude plugin marketplace add anthropics/claude-plugins-community claude plugin install next-steps@claude-community

🛠️⚡ Anthropic ปล่อยฟีเจอร์ "Mods" ใน Claude Code! ปรับแต่งการทำงานสาย Dev ได้ตามใจชอบ 🧵👇 1/ ปรับแต่งโหมดการทำงานได้ลึกขึ้น Anthropic อัปเดต Claude Code (เวอร์ชัน 2.1.287) เพิ่มระบบ Mods ให้เหล่านักพัฒนาสามารถเขียน TypeScript หรือ JavaScript สั้นๆ เพื่อปรับแต่งพฤติกรรม หน้าตา UI และการเรียกใช้ Tools ได้อย่างอิสระ 2/ ตัวอย่าง Mod เด็ดที่ชุมชนสร้างขึ้น 🚀 Mission Control: ติดตาม Progress การทำงานของ Code แบบ Real-time 🎨 Pixel-art Desks: แต่งธีมโต๊ะทำงานสไตล์พิกเซลแยกตาม Session 🛡️ Safe Delete: ฟีเจอร์ความปลอดภัย ป้องกันการลบไฟล์ผิดพลาดด้วยระบบ Undo 3/ เปลี่ยน Terminal ให้เป็น Cockpit ส่วนตัว นอกจากนี้ยังมี Mod ช่วยวาดไดอะแกรมอธิบายโค้ด ไปจนถึงตัวจัดการ Cache เพื่อยืดอายุ Session... อย่างไรก็ตาม เนื่องจาก Mods ได้รับสิทธิ์ในการเข้าถึงระบบสูง ผู้ใช้งานจึงควรเลือกดาวน์โหลดจากแหล่งที่น่าเชื่อถือเท่านั้น! ✨ #Claude

I made a mod for Claude Code: a better status line. Context, cache and rate limits as dot-matrix rings. What each turn cost. And the music you're playing, right above the prompt. Open source: https://github.com/Autumn1337/better-statusline

AI 写的落地页，一眼就能认出来：Inter 字体、紫到蓝渐变、卡片套卡片、彩色底上一行灰字。 今天 GitHub Trending 前排的设计插件 Impeccable，专治前端「AI 味」👇 是什么： · 给 Claude Code / Codex / Cursor / Gemini CLI 等 agent 用的设计 skill，Apache-2.0，约 7.7 万⭐，今天单日 +1,171 · 1 个 skill、24 条命令：/impeccable critique 设计评审，audit 查无障碍/性能/响应式，polish 上线前收尾，bolder/quieter 调浓淡，harden 补报错和边界态 · 61 条确定性检测规则，CLI 跑这部分不用 LLM、不用 API key 怎么试： 1. 先给官网体检：npx impeccable detect https://你的域名（扫渲染后的页面，本机要有 Chrome/Chromium/Edge；CLI 要求 Node ≥22.18）。我拿一个故意写烂的页面实测，报出 6 条：灰字对比度 1.3:1、Inter、紫色渐变、h1 直接跳 h4 等 2. 装进 agent：项目根目录 npx impeccable install，再在工具里跑 /impeccable init，它会写一份 PRODUCT.md 记下受众、用途和语气 3. Claude Code 也能 /plugin marketplace add pbakaus/impeccable；Codex 装完要去 /hooks 批准项目 hook 4. 放进 CI：npx impeccable detect --json .，有问题退出码为 2 注意：作者在 README 里写了，Claude Code 里装的 hook 不走模型工具审批，首次编辑会自动下载引擎二进制，无人值守跑之前先看一眼。 一句话：独立开发没设计师，至少别让用户第一眼就看出这页是 AI 糊的。 https://github.com/pbakaus/impeccable

A user-installed mod can approve a tool call. On a managed machine, that approval no longer beats a deny or ask on a nested part of a compound shell command. Claude Code 2.1.289, October 3, closed that gap. The rule has to match the inner command, not only the string you typed. A mod still runs inside the Claude Code process. It is not a second policy store. If the deny lives in managed settings, the mod does not get the last word. If it does not, do not read this release as a lock.

QA in progress 🔍 Build 44: "Minecraft Crossover" in GTA V × Minecraft is built. Now the AI tests it in game and an independent review checks it, live. Modded games, built and tested live by Claude Code.