Community posts

Claude Modsを使って 自作のエージェントとの電話機能作ったら快適すぎたww 『/realtime-voice』のコマンド打ったらあとは喋るだけでエージェントが勝手に仕事してくれる😂 ChatGPTのDotsだとモデルは結局GPTなので、Claudeモデルでやりたい俺はこっちだなw ごろ寝しながらでも仕事がどんどん進んで めっちゃ快適 コードもシェアしとくわ👇

Claude Codeに、プロンプトが届く前に書き換えられる拡張機能が入ったそうです。TypeScriptで数行から書けて、「rm -rfを止めて確認させる」「ツール権限を自動で承認・拒否する」といった制御をエージェントの動作パイプラインに直接差し込める。 こういう仕組みを「誰が・どのレイヤーで持つか」がエージェント時代のプロダクト設計の核心になってきますよね。 「Claude CodeにTypeScriptで「プロンプト書き換え」機能を追加できる拡張システム「Mod」登場 — AIエージェントの権限制御を開発者が直接制御可能に」 https://ai.freelance-job.com/article/6abee4db6f60ae186b878af6

DeepSeek Harness v0.2.1-alpha.1 ships an experimental Claude Code Mods compatibility layer, agent-created plugins, and built-in automated tasks. Are you upgrading? #DeepSeek #Harness #Claude #ClaudeCode #Mods https://open.substack.com/pub/aidisruption/p/deepseek-harness-v021-alpha1-released?r=2ajqea&utm_campaign=post-expanded-share&utm_medium=web

Claude Modsなんて使わなくても、CodexはOAuthをサードパーティクライアントで使わせてくれるが...？

Claude Code v2.1.287 ships Mods: JS/TS middleware intercepting tool calls inside the agent loop, with full user permissions, no review, no sandbox. 26.1% of 31,132 skills in one study already had vulns. The platform outran its trust layer.

Anthropic just shipped a pair programmer for your pair programmer "You Should Know" - a built-in Claude Code mod A side agent watches every turn and calls out mistakes on both sides: Claude → fake "done", skipped reads, swallowed errors You → risky pushes, leaked secrets, unread plans One command: /plugin enable cc-plugin-you-should-know@builtin Requires Claude Code 2.1.287+ with telemetry on A flag is only useful if something acts on it Copy the prompt that makes Opus 5.5 act on every one

From Claude Code 2.1.287 on, Mods are on by default. A mod is a small JavaScript or TypeScript file that stays loaded in the session. Its hooks form a middleware chain. A hook can watch an event, rewrite what the rest of the chain sees, or answer the event itself, and that is how a mod intercepts http://tool.call. It can draw UI, open a panel, and register a slash command. The guide’s examples are Token Weather, Blast Radius, and Replay Theater. Token Weather is a live forecast of the context window, drawn above the prompt, in about 80 lines. Blast Radius holds a Bash call such as rm -rf, git reset --hard, git clean, a force push, or a database migration, then opens a pane with Proceed and Cancel. Replay Theater records that turn’s Edit and Write calls and, on r or /replay, steps through the diffs in a pane. Some of Claude Code’s own features are mods, including AGENTS.md support and the /diff pane. The Decoder, not the guide, describes the built-in You Should Know plugin. It starts a separate agent that sends a Heads up when it spots something the user might have missed, and it is enabled with /plugin enable cc-plugin-you-should-know@builtin. Claude Code 2.1.287부터 Mods는 기본으로 켜져 있습니다. Mod는 세션 안에 유지되는 작은 JavaScript 또는 TypeScript 파일입니다. 훅은 미들웨어 사슬을 이룹니다. 훅은 이벤트를 지켜보거나, 뒤쪽 사슬이 보는 내용을 바꾸거나, 이벤트에 직접 답할 수 있고, 이 방식으로 http://tool.call을 가로챕니다. 화면을 그리고, 패널을 열고, 슬래시 명령을 등록할 수 있습니다. 안내의 예시는 Token Weather, Blast Radius, Replay Theater입니다. Token Weather는 프롬프트 위에 그리는 컨텍스트 창의 실시간 예보로, 약 80줄입니다. Blast Radius는 rm -rf, git reset --hard, git clean, 강제 푸시, 데이터베이스 마이그레이션 같은 Bash 호출을 붙잡아 두고 Proceed와 Cancel이 있는 패널을 엽니다. Replay Theater는 그 턴의 Edit와 Write를 기록하고, r 또는 /replay에서 그 diff를 패널에 단계적으로 보여 줍니다. AGENTS.md 지원과 /diff 패널을 포함한 Claude Code의 일부 자체 기능도 Mod입니다. 안내가 아니라 The Decoder가 내장 플러그인 You Should Know를 설명합니다. 사용자가 놓쳤을 수 있는 것을 보면 Heads up을 보내는 별도 에이전트를 띄우며, /plugin enable cc-plugin-you-should-know@builtin 으로 켭니다. https://claude.dev/blog/getting-started-with-claude-code-mods/ https://github.com/anthropics/claude-code

貼り付けた画像が、 やっと見える。 Claude Code の入力欄の上に、 サムネイルを出す mod が 公開されました🖼️　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　　 名前は image-viewer。 これまでは、画像を貼っても「[Image #1]」という文字のタグが出るだけでした。 入れると、貼った瞬間に入力欄の上へサムネイルが出ます。 ✅ 縦横比は保たれる ✅ 送信するか、タグを消すと消える ✅ 手元だけで動き、通信はしない 入れ方は、Claude Code の中で3つのコマンドです。 マーケットプレイスの追加、インストール、再読み込みの順です。 条件があります⚙️ Claude Code のバージョン2.1.287以降、macOS か Linux、そして Ghostty や kitty のように、画像の表示方式（kitty graphics protocol）に対応した端末。 ほかの端末では、今までどおり文字のタグのままです。 Claude のデスクトップアプリには、最初から画像のプレビューがあるので、この mod は何も描きません。 使っている端末が条件に合うかを、先に確かめておきたいところです。 https://x.com/jarrodwatts/status/2106153410697564235

Claude Code 2.1.287で「Mods」という新機能が追加された。プラグイン経由でClaude Codeの挙動やUIをカスタマイズできる仕組みだが、セキュリティ面で押さえておきたい注意点が複数あるので整理する。 1. Modsはサンドボックス化されていない 公式ドキュメントには、Modは自分(ユーザー)の権限で動作するコードであり、ファイルの読み書き、プロセスの起動、ネットワーク通信ができると明記されている。実行環境の分離はされないため、環境変数に置いているAPIキーなども読み取り得る権限を持つ。信頼できる作者・マーケットプレイス以外のものは入れない方がいい。 2. 導入は/pluginコマンド経由 Modはプラグインの一種としてマーケットプレイスから配布され、/plugin install <name>@<marketplace>で導入する形。実はClaude Code自身の一部機能(diff、sec-default、telemetry、agents-mdなど)もMod実装になっていて、/pluginコマンドのInstalledタブでBuilt-inとして一覧確認できる。 3. バージョン要件はv2.1.287以降 この機能を使うにはClaude Codeをv2.1.287以降にアップデートする必要がある。/statusなどで手元のバージョンを確認してから試すこと。 4. 実務での注意 Sandboxモードを有効にしてもシェルコマンドの実行環境が隔離されるだけで、Mod自体はその隔離の対象外になる。チームや組織で導入ルールを決める場合は、「誰が」「どのマーケットプレイスの」Modを承認するかを事前に決めておいた方がいい。 元ネタは@ClaudeDevsの告知(2026-10-01〜10-02)。便利な機能だが、インストールするコードを増やすという感覚で扱うのが安全。 出典: https://code.claude.com/docs/en/plugins/mods/overview 出典: https://x.com/ClaudeDevs

I packaged exactly this setup as an open-source Claude Code plugin: Sonnet workers/researcher/explorer, Opus escalation, a verifier before "done", plus a local dashboard showing what each agent is doing. https://github.com/emiirhandemirci/SubDeck

Claude Code mod to remind you of next prayer

Claude Code を使うついでに、外国語の読む・書く・聞く・話すを練習できるmod「claudelingo」をリリースしました！ 普段やり取りをしている中で「これ本当は英語でやったらすげえ勉強になるのにな、LLM相手だから気兼ねないし」と思ってました。 が、めんどいし、そもそもなんて言っていいのかわからない、そもそも仕事だったらそっちの早さが優先されるとか、わざわざやらないです。 って時に、本流チャットを一切邪魔せず、サイドで 段階的に 学習できるようにしました。 コアイメージの図解の動的生成や入力欄でのリアルタイム校正とか LLMならではの学習方法も盛り盛りにしてます claude plugin marketplace add yodakeisuke/claudelingo claude plugin install claudelingo@claudelingo

【改造の工夫】待ち時間の使い方に差が出ます！ Claude CodeをModsで拡張し、AIが作業している間に本人はDoomで対戦していたという海外の投稿が話題になっていました🔍 ここからが中小企業にとっての本題です。AIに任せている時間をどう使うかは、社内ではまだ決まっていないことが多いです。 ・AIが動いている間、何をするか決めていない ・出てきた結果の確認を後回しにして忘れる ・待ち時間の使い方を相談できる相手が社内にいない 正直、ここは工夫のしどころです。待ち時間を空費にするか次の判断に使えるかは、相談できる相手がいるかどうかで変わります！📝

DeepSeek Harness adds experimental Claude Code Mods compatibility layer https://aiunderstanding.org/news/deepseek-harness-adds-experimental-claude-code-mods-compatibility-layer?utm_source=x&utm_medium=social&utm_campaign=breaking_ai_news

Claude Code now has mods. I built one for PaF (phone-a-friend) Ask Codex or another model for a second opinion and watch it happen: the question, the reviewer's progress, then its findings in its own words, one thread per branch, round after round. A few other things since April: • Works from Codex, OpenCode and pi, too • New backends: Antigravity and pi • Every review is saved as a task you can find again later https://github.com/freibergergarcia/phone-a-friend

Claude Codeに「Mods」が入った。 Thariq（@trq212）が紹介していたので、公式ドキュメントと手元のCLIで全部裏を取った。 一部、ポストの説明より踏み込める。 元記事→ https://x.com/trq212/status/2105734197801562264

Claude CodeでMODが扱えるようになったので、プロンプト枠の上で一緒にカタカタするフトアゴのキャラクター作ってみた (*‘ω‘ *)カタカタ♪

Claude Codeの待ち時間にDoom対戦。Modsの使い方が自由すぎる。

## Claude Code 最新版更新 Mods 能力，致敬 DeepSeek Harness？ Claude Code v2.1.287 悄悄加了个叫 Claude Mods 的东西，本质是往进程内的事件流里挂 TypeScript/JavaScript 函数，能做三件事：观察、改写、接管。比如每次工具调用后记账，或者把危险命令参数偷偷换成安全选项，甚至直接跳过默认逻辑自己返回结果。 最有意思的是 Anthropic 自己在拿它做内置功能迁移。现在 /diff 面板对应 cc-plugin-diff，加载 AGENTS.md 对应 cc-plugin-agents-md，官方还说要慢慢把更多内置能力拆成 Mods，让宿主缩成一个 Core。这思路跟 DeepSeek Harness 的“一切皆插件”撞上了，不过两边路子不同：dsh 是启动时按 Profile 拼整棵 Cordis 插件树、换 AgentLoop 和模型 Provider；Mods 是在跑起来的宿主上开事件中间件，管的是单轮模型路由和工具执行。 但有个细节我挺在意，官方写了句 Mods aren't sandboxed。Mod 里用 $.fs 或 $.process 时，行使的就是你当前用户的系统权限，能读配置文件里的 API Key。而且 Bash 沙箱管的是模型自己发的命令，Mod 启动的外部进程在沙箱之外。装第三方 Mod 真得留个心眼。 #ClaudeCode #AI编程 #DeepSeek

DeepSeek's own harness added a Claude Code Mods compat layer about 37 hours after Anthropic shipped Mods. The changelog is blunt: the goal is to prove the Mods API is "roughly a subset" of what DeepSeek Harness plugins already do. Not full compatibility for your mods. The model was never the moat. The plugin API is.

Morning Update OpenAI releases GPT-6 Astra, its new flagship model for coding, cybersecurity and science OpenAI has introduced GPT-6 Astra, a new generation of its AI models. The company positions it as its most capable and safety-aligned model yet, with stronger performance for using computers, writing software, cybersecurity work and scientific tasks. The launch also makes clear that OpenAI’s public product line has moved beyond the GPT-5 generation, even as speculation about later models such as GPT-7 and GPT-8 remains unconfirmed. Google DeepMind unveils a new way for developers to build longer-running AI agents Google DeepMind presented an Interactions API and a Managed Agents offering for Gemini-based software. The tools are designed to let developers keep an agent’s task history and tool use on the service side, rather than repeatedly rebuilding context in every request. The aim is to make agents that use search, web pages and company tools easier to operate across multi-step jobs. Anthropic’s Claude Code adds “Mods” for changing how the coding agent works Anthropic has introduced Mods for Claude Code, allowing developers to customize the coding assistant’s behavior and interface from inside the product. Unlike external add-ons that merely connect extra tools, Mods are intended to alter the agent’s working environment itself. That could enable more tailored coding workflows, while also increasing the need for teams to review third-party extensions before installing them. Want to find out more about any of these topics? Search them on http://FoundOnArdy.com.

Karpathy: we'll spend more time understanding LLM output. I built a Claude Code mod for it: Claude picks the fastest format to read (plain STE-style text, diagram, throwaway HTML page, video) based on your project. /understand to force one. Free: https://github.com/elharchaoui/claude-mods

Claude Code로 자동화 돌려두고, 매일 "오늘 거 돌았어?" 물어보고 계신 분들.. 물을 때마다 Claude가 로그 열어보고, 기록 찾아보고, 그 답을 기다리게 되지 않나요? --> 이번 주 Claude Code에 생긴 "모드(Mods)"로 이걸 화면에 박아둘 수 있습니다. 예를 들면 입력창 바로 위에 이런 한 줄이 늘 떠 있게 하는 거예요. "📊 오늘 예약 작업 3개 중 3개 완료 · 실패 0 · 마지막 실행 09:00" (예시 문구입니다. 무엇을 띄울지는 내가 정해요.) 1분마다 다시 읽게 해두면 상태가 바뀔 때 알아서 바뀌고, 자세히 보고 싶을 땐 "/today" 같은 명령 하나로 바로 볼 수 있습니다. 오해하실까봐 먼저 말씀드리면.. 스킬이나 MCP는 Claude가 "무엇을 할 줄 아는지"를 늘려줍니다. 결과는 대화창에 글로 나오죠. 모드는 Claude Code "화면 자체"에 부품을 붙입니다. 입력창 위 한 줄, 옆 패널, 진행·취소 버튼 같은 것들이요. 쉽게 말하면 비서한테 매번 묻던 걸, 계기판 눈금으로 하나 달아두는 겁니다. 그리고 이거 코드 몰라도 됩니다. 원하는 모드를 말로 설명하면 Claude가 써 준다고 공식 문서가 안내하는데요, 저도 이런 식으로 말해서 하나 만들어 봤습니다. "매일 돌리는 작업 결과를 입력창 위에 한 줄로 띄우고, /today 치면 자세히 보여주는 모드 만들어줘" 이런 명령은 Claude를 거치지 않아서 기다림이 없습니다. Claude가 다른 일 하는 중에도 바로 쳐볼 수 있어요. 그러면 이걸 어떻게 써먹냐? --> 매일 확인하던 상태를 한 줄로. 예약 작업, 봇, 콘텐츠 파이프라인 돌리시는 분들 --> 자주 묻던 질문을 "/명령어" 하나로. 답이 바로 나옵니다 --> 실수를 코드로 막기. 공식 샘플 "blast-radius"는 삭제나 강제 업로드 같은 위험한 명령 직전에 진행·취소 버튼을 띄웁니다 (메모리에 적어둔 규칙은 Claude가 잊을 수 있는데, 모드는 실행 직전에 붙잡는다는 게 차이예요.) 한 가지는 꼭 짚고 갈게요. 모드는 내 권한으로 도는 코드이고 샌드박스가 없습니다. 파일 읽기·쓰기, 프로그램 실행, 입력한 프롬프트까지 볼 수 있다고 공식 문서가 경고합니다. 남이 만든 건 믿을 수 있는 곳에서 온 것만 설치하세요. (Claude Code 2.1.287 이상, 터미널과 데스크톱 Code 탭에서 보입니다. VS Code 확장에서는 안 보이고, 회사 계정은 관리자가 막아뒀을 수 있어요.) (Claude가 만든 모드는 그 세션에서만 켜집니다. 계속 쓰려면 폴더를 따로 두고 "claude --plugin-dir 폴더"로 열면 돼요.) 나온 지 며칠 된 기능이라 문서도 바뀔 수 있다고 적어뒀습니다. 2026년 10월 기준이에요. Claude Code로 뭔가를 매일 돌리면서 같은 질문을 반복하는 게 고민이라면, 이번 주 안에 그 질문 하나를 화면 한 줄로 바꿔보는 걸 목표로 잡아보는 건 어떨까요? 만드는 법: https://code.claude.com/docs/en/plugins/mods/create

Someone show me a Claude Code mod that saves you ten clicks every day. Tiny demo, what it replaces, and how you remove it. That’s enough to get me interested.

Claude Code's first built-in mod is the interesting half of Mods for me: "You should know" (cc-plugin-you-should-know@builtin). side agent watches the session and fires a Heads up when it spots something you or the primary Claude might skip. enable with /plugin enable cc-plugin-you-should-know@builtin; needs first-party session + telemetry on in the current docs. builder take: an oversight agent is useful. an unsandboxed mod channel that can also approve http://tool.call is still a supply-chain surface. turn on the watcher, keep marketplace installs on an allowlist, and treat Mods like local binaries.

10 月 1 日,Anthropic 发布 Claude Code mods。 10 月 2 日上午,DeepSeek Harness 就提交了一个「Claude Code mods 兼容桥」。 但翻开 DeepSeek 自己的兼容性文档:官方 4 个内置 mod,一个都跑不了。 能跑,不等于能放心跑 🧵 https://x.com/i/article/2106673977766649856

Claude CodeのCompaction時間を52秒→0.26秒にして、要約せずロスレスにするpluginを作りました。爆速compactionを試して欲しいです https://github.com/yottayoshida/lossless-compaction

🔍 How to view DeepSeek Harness adding Claude Code Mods compatibility in its latest version? Zhihu contributor & DeepSeek Harness team member 崔添翼 @tianyi shares his thoughts firsthand, explaining the purpose and thinking behind it: Thanks to everyone for your interest in DeepSeek Harness. I’m 崔添翼 from the DeepSeek Harness team, and I’m the author of the Claude Code Mods compatibility layer added in v0.2.1-alpha.1. First, a disclaimer: the “Claude Code Mods compatibility layer” in DeepSeek Harness is an experimental feature in the alpha version. There are still many areas that need improvement, and we cannot guarantee 100% compatibility. The current compatibility documentation lists the details. 🔗：https://github.com/deepseek-ai/deepseek-harness/blob/dsh-v0.2.1-alpha.1/docs/subsystems/claude-code-mods.zh.md The main purpose of building this compatibility layer is to verify whether the extension capabilities provided by Claude Code Mods to plugin authors are, broadly speaking, a subset of what the “everything is a plugin” architecture of DSH provides. It is not yet possible to run all Claude Code Mods seamlessly in DeepSeek Harness. If this is considered a worthwhile goal, it may be technically possible in the future. The core philosophy of DeepSeek Harness is “everything is a plugin.” This principle was established from the very beginning of the project, before we wrote the first line of code, just as the philosophy of “open source” was established when DeepSeek was founded, before we released our first model. Open source is our original intention, not something we were forced into. Making our Agent Harness product as extensible as possible through an open and open-source approach is also our original intention—not something we adopted because we were forced to or because others did it. What is the use and significance of an Agent Harness where “everything is a plugin”? In my view, “everything is a plugin” has at least three meanings: First, we hope to meet diverse user needs through plugins and the community, so that everyone can use an Agent Harness truly built for themselves and suited to their needs. By training models to write plugins and modify the Agent itself, we hope Agent products can eventually change according to what users need and want—even reaching the effect of “what you say becomes what happens.” The “Creation Mode” is already a small step in this direction: it allows the Agent to write and install plugins based on user needs, extending the Harness’s capabilities and interface. As a new concept, “Creation Mode” is certainly still rough and needs refinement, so we welcome everyone to try it and share feedback. Second, we hope to meet the needs of enterprises in different industries for deploying Agent products through a plugin-based approach. Enterprises across industries, as well as FDE-type teams helping them implement AI, can use plugins to accumulate industry expertise and connect industry infrastructure, without repeatedly reinventing the wheel. This allows them to use cutting-edge model capabilities and open-source Agent Harness frameworks more efficiently and at lower cost, ultimately improving enterprise productivity through AI. We have already seen examples of DeepSeek Harness being deployed in some industries: developers have built DSH plugins for specific industries to create “AI workbenches.” These industry-specific DSH plugins are generally not publicly open-sourced online. This is compliant with the MIT open-source license and is something we support and recognize. Third, we hope to use the “everything is a plugin” Agent Harness architecture to advance the idea of “Agent self-evolution.” With a powerful plugin system, we can provide the flexibility to change any part of how an Agent works, allowing the Agent to modify itself through plugins while running, and ultimately raise its intelligence ceiling. This direction requires the joint evolution of models and Harness, and the related work has only just begun. There is still a long road ahead. From an engineering perspective, we need to more fully implement hot-plugging and hot-reloading of plugins. Cordis provides a solid theoretical and engineering foundation, but there are still many engineering challenges to solve. From the perspective of model research, we need to train large language models in new ways, so that models know they can modify their own Harness, learn how to iteratively improve and evolve themselves by changing their Harness, and ultimately raise their intelligence ceiling. Therefore, the DeepSeek Harness team is actively recruiting researchers who understand both large-model research and training, and who are interested in improving Agent intelligence through the joint evolution of models and Harness, gradually moving toward the vision of “Agent self-evolution.” I personally spend a considerable part of my working time recruiting and interviewing for DeepSeek. Many candidates ask me: What is the goal of the DeepSeek Harness team? My answer is: DeepSeek’s goal is to achieve AGI. The goal of the DeepSeek Harness team is to advance AGI from the perspective of Harness. So what counts as "advancing AGI from the Harness angle"? Colleagues can each have their own view; mine is that only innovation truly advances AGI. If we copied a Claude Code or a Codex — a product basically the same as other Agent Harnesses, from technical philosophy to user experience — that would not advance AGI. Users can already run DeepSeek models in Claude Code, Codex and others. If ours is about the same as those, we deliver no incremental experience, and it isn't our way of advancing AGI. Therefore, advancing AGI must be about innovation. This Claude Code Mods compatibility experiment is one concrete attempt to test the extensibility of the DSH plugin system through an actual implementation. At the DeepSeek Harness team, we hope to create world-class innovations—work that shapes how everyone sees and defines the Agent Harness, and that the world's AI giants will recognize and follow. From both the team and product perspectives, DeepSeek Harness is still very young and unpolished. We welcome your suggestions. The road ahead is long, and AGI is far from its endgame.

AIに仕事を任せるほど、「何を見落としたか」が見えなくなってませんか？modの使い方が公式からできました。個人的には必須modだと思う。 Claude Codeの新機能「You should know」は、その盲点を拾うための組み込みmod。 仕組みは、作業するClaudeの横で別のエージェントが動き、Claudeの出力を観察する。ユーザーやClaudeが見落としそうな重要情報を見つけると、入力欄の上にメモを表示して知らせる。長いタスクを任せる場面を想定した機能だ。 ここが面白い。 作業を進める役と、気づくべきことを拾う役を分けている。 AIが高速に作業すると、人間は途中経過を追い切れなくなる。「説明されていた」と「理解していた」の間に、穴が開く。You should knowは、その穴を埋める補助になる。 この機能の価値は、「AIがやってくれる範囲」だけでなく、「任せた人が状況を把握できる範囲」に目を向けたことにあると思う。 試すには、Claude Code内で以下を実行。 /plugin enable cc-plugin-you-should-know@builtin v2.1.287で追加。初期状態では無効で、公式の変更履歴では「テレメトリが有効なファーストパーティセッション」が対象。組織で利用可能なら、/plugin → Installed → Show disabledからも確認できる。 見落としを全部検出する保証はない。それでも、AIに長い仕事を任せる人ほど、この「横で気づく役」は試す価値がある。 作業速度が上がっても、判断に必要な情報まで流れていっては困る。 出典：ClaudeDevs公式発表 https://x.com/ClaudeDevs/status/2106118517447876618 公式ドキュメント： https://code.claude.com/docs/en/plugins/mods/overview 関連記事：Addy Osmani「Getting started with Claude Code mods」 https://claude.dev/blog/getting-started-with-claude-code-mods/

UPDATE: Claude was able to reverse engineer all the expansions and the later Smash Frenzy 2 rebadge release. Turns out later builds of the Magic Ball 2 engine can just support as many level packs as you please once unpacked and DRM free. It's EXTREMELY easy to mod. So here's Magic Ball 2: Community Edition Includes MB2: New Worlds and the base game merged together. https://kippykip.com/threads/pc-magic-ball-2-community-edition-restored-for-windows-10.5797/