Community posts

THIS GUY SHIPPED A DONE / NOT-DONE VERDICT JUST SO "ALL DONE?" STOPS BEING A GUESS Open Steps: plain-language skills — done or not, debt left, safe to close, next step that needs you. MIT · Claude Code plugin · https://github.com/kharmanskyi/open-steps

@Maxmarinex @RockytheGoodBoy @keremyilmazyt @sefacarrier @forbir1 @SismanlarTV @YabguNizamPasha @apocaselips @talentlenus @kanfyt @MuratEnginEkin @YKayaalp07 @NeOlurduYT Civilization 5'e mod geliştirmeye başladım. İlerleyen haftalarda paylaşmak nasip olur umarım. Paradox'a muhtaç değiliz. Elimizin altında claude code var. Paradox ne yaptığının hala farkında değil.

Claude Code Mod制作まで行けるとは思わんやん

If you run Claude Code on a wide monitor, set maxProseWidth in your settings (v2.1.282, 40 minimum). Replies wrap at that column while code blocks and tables stay full width. Also useful for extra padding between content and any mod window that is open at the same time.

love the claude code mods feature, definitely opening the door for extensibility i built an example for being able to have github prs/issues accessible through just typing #github https://github.com/jabreeflor/gh-refs-mod

Claude と 17 日間、Drag'n Wash の Mod を作ったら、どれくらいかかったのか。会話の記録を全部数えてみました。 84 億トークン、API の定価なら約 66 万円（実際は月額プラン）。 外注したら約 8,000 万円…… 同人誌に換算すると、TRC がほぼ埋まりました（笑）

5 HOTTEST AI GITHUB REPOS RIGHT NOW: 1. deepseek-ai/deepseek-harness — ~238k stars — Plugin-first coding agent harness https://github.com/deepseek-ai/deepseek-harness 2. anomalyco/opencode — ~210k stars — Open-source terminal coding agent https://github.com/anomalyco/opencode 3. anthropics/claude-code — ~148k stars — Anthropic’s terminal coding agent https://github.com/anthropics/claude-code 4. openai/codex — ~127k stars — OpenAI’s lightweight terminal coding agent https://github.com/openai/codex 5. stablyai/orca — ~79k stars — ADE for running fleets of coding agents in parallel https://github.com/stablyai/orca

Taking a step away from sitting on Claude to ask why the hell are people even doing these 3D mods when it always and I mean always looks worse HD2D games look great because they had intent behind their styles and spritework, it was intentional. This is just ugly.

Built a @herdrdev plugin that picks Claude Code, Codex or @opencode + model + effort for a task, then launches it in a new pane. Routing runs locally with Laya in ~150 ms. No API key, your task text stays on your machine. Inspired by agent-router. https://github.com/candidosales/herdr-laya-router

Made a plugin from parts of my Claude Code setup (currently in review) plus a few new ideas I'm trying out right now lmk what you think: https://github.com/simon-init/cguard

Claude Code's early-access function hooks are quite fun. I built a plugin to add more personality to Claude Code https://claudecafe.dev/plugin

https://zenn.dev/nogu66/articles/claude-code-function-hooks-claude-mods Claude Mod、試しにやってみたら地味に結構いいかも。 auto compactだと時間ベースで制御できないから中座したときに再読み込みが発生しがち。Xで見かけたアイデアだけ参考に50分経って入力がないときは自動でcompactが走るようにしてみた。 再読み込みがどれくらいUsage消費してるかはちゃんと調べたわけではないが、ロングセッションがUsageの大部分を占めているの把握してるので結果がよければ共有するかも

AWSの構成図、Claude Codeに箇条書きで構成を渡すだけでhttp://draw.ioの図が簡単にできちゃう、、SEからしたらうれしすぎる AWS公式アイコンで描いてくれて、あとからhttp://draw.ioで普通に直せる Opus 5.5で1枚〔3分半〕くらい 入れ方（Claude Codeで2行） /plugin marketplace add sagochiko/aws-drawio-diagram-skill /plugin install aws-drawio-diagram@aws-drawio-diagram-skill 頼み方は「以下のAWS構成の構成図を作って」＋箇条書きだけ 枠の入れ子や線の引き方のルールはSkill側に書いてあるから、構成だけ書けばいい 既存の.drawioに「WAFを足して」みたいに頼むと、描き直さずにそのファイルに足してくれる モデルはOpus 5.5がいい。作者の計測だとSonnet 5でも描けるけど、直しの往復が増えて時間は3倍くらいかかる 苦手なのは、マルチリージョン、オンプレとのハイブリッド、30ノードを超える構成。線の経路は描くたびに少し変わる ※図は架空の構成で試したもの #ClaudeCode #AWS

明日からの業務に向けて、Claude Mods に入門しませんか？

今日爆款 Hermes 插件🔥 Hermes 跨端记忆金库 + 可删改召回 + 边干活边说话 + 独立邮箱工牌 + 一句话造 Bot 五连杀！ 全网玩家把 Hermes 玩成了下一代「记忆能搬家 + 记错能改 + 嘴和手同时在线 + 邮件有工号 + 角色能量产」： 1️⃣ Signet-AI/signetai（https://github.com/Signet-AI/signetai） 跨 Hermes / Claude Code / Codex / Pi / OpenClaw 同步记忆、身份文件、会话誊本和密钥。 “第二大脑第一次能换工装，而不是每个 harness 各自失忆一次”！ 2️⃣ 410979729/scope-recall-hermes（https://github.com/410979729/scope-recall-hermes） Hermes + Codex 长期记忆：本地存、带出处召回，能查、能改、能删。 “记住的东西第一次有账本，而不是一笔糊涂账写进黑洞”！ 3️⃣ bielcarpi/hermes-live-voice（https://github.com/bielcarpi/hermes-live-voice） 自托管实时语音网关：你继续说，Hermes 后台继续干，本地 / Gemini Live / OpenAI Realtime 三路。 “第一次真正边聊边施工，而不是说完一句等它写完一篇检讨”！ 4️⃣ openmailsh/hermes-plugin（https://github.com/openmailsh/hermes-plugin） 给 Agent 一张独立邮箱工牌，收发走 OpenMail，不借用老板私人邮箱。 “Agent 第一次有自己的收件箱，而不是偷偷用你的 Gmail 回客户”！ 5️⃣ BkashJEE/hermes-bot-forge（https://github.com/BkashJEE/hermes-bot-forge） 一句话「给我做一个某某角色」，直接出名字、脸、SOUL.md、记忆、工具、技能、例行和网关。 “Bot 第一次是工厂件，而不是周末手搓一份人格散文”！ // 这波直接把 Hermes 从「会自我进化的单兵」推到「记忆可搬运、召回可审计、语音不断档、邮件有工号、角色能量产」的完整操作系统。今晚装，明天舰队和收件箱一起起飞。

What's new in Claude Code 2.1.283 system prompts (+11,976 tokens): Expands attached-machine routing and failure handling, adds managed model allow-and-deny controls, hardens self-hosted runner credential and certificate guidance, and introduces plugin authoring, background-task inspection, and safer Artifact publishing rules. - NEW: Agent Prompt: Security monitor attached machine call results—Treats output from calls served by a user-attached machine as private data; sending it externally is judged under Data Exfiltration. - NEW: Data: availableModelsMatch setting—Managed setting choosing prefix or exact matching for availableModels; exact mode stops a model ID from allowing later releases until they are listed. - NEW: Data: deniedModels setting—Managed setting blocking models even when availableModels allows them; a model ID blocks every spelling of that version, and Default steps down past blocked models. - NEW: Data: Published model catalog seed guidance—Internal note on the compiled-in model catalog copy used before the first fetch and as the version floor; rows must not be hand-edited. - NEW: Data: SDK set max thinking tokens request schema—Documents resetting versus keeping the thinking budget and a session-scoped thinking_display override, including when highlights is accepted, downgraded, or refused. - NEW: Data: SDK system init plugin_errors field—Lists plugins that failed or only partially loaded; Remote Control workers always omit the key, so its absence does not prove a clean load. - NEW: Data: Self-hosted runner Anthropic git proxy credential warning—Warns that sources on ungoverned hosts need credentials outside the HOME-level git config --use-anthropic-git-proxy replaces, or private clones fail. - NEW: Data: Self-hosted runner client certificate relay warning—Warns that runner-wide TLS client certificates would be offered to the session relay, not governed hosts; suggests direct listing, unsetting, or per-URL scoping. - NEW: Data: Self-hosted runner GIT_ASKPASS governed hosts warning—Warns that a global askpass program could send this machine's credential through Anthropic's relay; directs scoping or unsetting it, core.askPass, and SSH_ASKPASS. - NEW: Data: Self-hosted runner GIT_SSL_CAINFO trust bundle warning—Explains that GIT_SSL_CAINFO prevented building the combined certificate file for Anthropic-managed git, and recommends a per-server sslCAInfo entry instead. - NEW: Data: Self-hosted runner GIT_SSL_NO_VERIFY lifecycle hook exception and session relocation note—Explain that inside sessions, and in hooks of sessions using Anthropic-managed git, the variable becomes http.sslVerify=false while the git mount stays certificate-checked. - NEW: Skill: /doctor prompt-audit configuration scope—Scopes /doctor prompt-audit to Claude Code configuration loaded in this project, skips settings files and secrets, and treats audited files as data, not instructions. - NEW: Skill: Plugin authoring—Guides writing hot-reloading function-hook "mods" such as panes, status lines, toasts, and tool-call hooks; the user enables hot-reloading once per session. - NEW: System Reminder: Attached machine stopped answering—Marks a command's outcome unknown when an attached machine stops answering; forbids non-idempotent retries and further calls to it this turn. - NEW: System Reminder: Attached machine untrusted attachments refusal—Explains that calls to an attached machine are refused while untrusted repositories or files are attached, and how the person can clear or avoid the block. - NEW: System Reminder: Directory sync file store exhausted—Warns that the session file store, or this environment's share of it, is used up, so further changes no longer sync to the user's machine. - NEW: System Reminder: No attached machine request guidance—Tells cloud sessions without an attached machine when to request the user's computer, how to handle offline or unanswered machines, and what to keep doing locally. - NEW: System Reminder: Remote machine-only resources routing—Sends tasks needing machine-only resources (platform tools, devices, logins, internal hosts) straight to the attached machine, but not project build failures or blocked public sites. - NEW: System Reminder: Unreachable attached machines—Directs finishing all other work here while attached machines are unreachable, reporting what waits, and retrying once only when the user asks, never polling. - NEW: Tool Description: Artifact stale publish saved-source guidance—Refuses publishes not built on the live Artifact version, pointing to its saved full source and requiring edits merged onto it rather than rebuilt from memory. - NEW: Tool Description: Bash (attached machines)—Explains routing individual commands to a user-attached machine with a per-call machine field, going there directly for machine-only needs, and never calling offline machines. - NEW: Tool Description: GetTask—Reads the state of a background Bash command by task ID; notes Claude Code often calls it automatically and forbids using it to wait. - NEW: Tool Parameter: Artifact preview action—Parameter description for the check tool's preview action, with optional viewport widths and themes, height-capped screenshots, and a layout/load checklist. - REMOVED: System Reminder: Directory sync restore up to last completed turn—Drops the notice that recovery reached the end of the last completed turn; the remaining restore notices now describe recovery by upload instead. - Data: Artifact runtime capability declarations—Adds that a page republishing itself through the artifact capability must send its whole document in the Artifact tool's exact skeleton shape. - Data: Managed Agents outcomes—Rubric uploads now use the non-beta client.files.upload(...), and deliverables note the managed-agents-2026-04-01 header files.list needs for scope_id. - System Reminder: Directory sync disabled after initial checkout failure—Can now add that some of the agent's files, moved into a trash folder to make room, could not be moved back and remain there. - System Reminders: Directory sync full and partial environment restore—Restored work now dates from one of the earlier environment's uploads rather than a turn boundary; changes made after that upload are missing. - Tool Description: Artifact type file-backed content update guidance—Once the Artifact's own files have been seen, files written or read earlier count as current until a publish is refused; refusals must be followed. - Tool Description: New file-backed Artifact type content guidance—For pinned content, later edits skip rereading files already seen, may publish several changed files in one call, and must follow any publish refusal. Details: https://github.com/Piebald-AI/claude-code-system-prompts/releases/tag/v2.1.283

this week in open source ai: 23 new projects added, Sep 14-Sep 20, 2026 (Mon-Sun UTC). 1. Colibri: Zero-dependency C inference engine that runs large Mixture-of-Experts models locally across disk, RAM, and VRAM. https://github.com/JustVugg/colibri 2. Claude Code Local: MLX-native Anthropic Messages API server for running the unmodified Claude Code CLI with local Apple Silicon models. https://github.com/nicedreamzapp/claude-code-local 3. nemotron-omni-mlx: MLX runtime for Nemotron 3 Nano Omni vision and audio towers on Apple Silicon, tested against NVIDIA’s PyTorch reference. https://github.com/nicedreamzapp/nemotron-omni-mlx 4. AISIX: Self-hosted Rust AI gateway with OpenAI and Anthropic-compatible APIs, routing, traffic policies, metrics, and trace export. https://github.com/api7/aisix 5. BrowserSkill (Tencent): CLI and extension that let AI coding agents control authenticated browser sessions in background tabs. https://github.com/Tencent/BrowserSkill 6. Ordewell: Terminal planner that turns a goal into ordered, editable coding-agent tasks with per-task runner, model, and mode. https://github.com/ordewell/ordewell 7. DeskcommCRM: Self-hosted sales CRM with autonomous AI agents, WhatsApp, tenant RAG, and MCP tooling. https://github.com/melgarafael/DeskcommCRM 8. OpenResearch: Local-first workspace and harness for coding-agent research, literature review, hypothesis generation, and experiments. https://github.com/alphaXiv/OpenResearch 9. WeKnora: Knowledge platform combining RAG, reasoning agents, wiki generation, and multi-source document ingestion. https://github.com/Tencent/WeKnora 10. YuE: Full-song music generation model with melody and chord planning, acoustic synthesis, zero-shot covers, and editing. https://github.com/multimodal-art-projection/YuE 11. VoiceStudio: Local-first voice platform for multi-engine text-to-speech, voice cloning, dubbing, and dictation. https://github.com/debpalash/VoiceStudio 12. Thursday: On-device voice assistant combining a live conversation with background model agents, shell, browser, skills, and MCP. https://github.com/cgoinglove/thursday 13. Octop (Tencent Cloud): Self-hosted multi-user AI assistant platform with long-term memory, MCP tools, and sandboxed execution. https://github.com/TencentCloud/Octop 14. ENZO: Self-hosted AI workspace for agents, skills, and tools, using provider API keys sealed in the browser. https://github.com/theguysudo/ENZO 15. Clips Kitty: Windows desktop app that turns long videos and livestreams into vertical clips with captions and local highlight selection. https://github.com/ColinGPT9/clips-studio 16. Voz: macOS dictation app that types speech at the cursor offline, with on-device meeting recording and summaries. https://github.com/quietbin/voz-mac 17. Codex-X: Cross-platform management app for Codex Desktop and CLI, with provider switching, failover, history, MCP, and usage analytics. https://github.com/yynxxxxx/Codex-X 18. Atlas: Local-first agentic development environment connecting Git commits with agent sessions, tool calls, traces, and shared memory. https://github.com/pacifio/atlas 19. jev-use: Claude Code, Codex, and pi plugin routing no-output agent-loop steps to a typed judgment model. https://github.com/shitianfang/jev-use 20. GitDiagram: Turns GitHub repositories into interactive architecture diagrams with source links and Mermaid export. https://github.com/ahmedkhaleel2004/gitdiagram 21. MulmoTerminal: Browser grid for live Claude Code and Codex sessions, with PTYs, status, persistence, and a Git worktree per cell. https://github.com/receptron/mulmoterminal 22. Worktrunk: CLI for Git worktree management and automation of parallel AI coding-agent workflows. https://github.com/max-sixty/worktrunk 23. Webcmd: Self-learning browser infrastructure that compiles site navigation into deterministic per-site CLI commands. https://github.com/agentrhq/webcmd explore the full open-source ai list at https://awesomeosai.com/

Play chess with Jev inside Claude Code The Claude Code Chess Mod now supports your @typesafeai API key, so you can play against Jev instead of Claude I can tell you the response is almost instant xD, so these games are going to be short https://aitmpl.com/component/mod/games/chess

Jev Skill Suggestion just passed 1.2k downloads Trending: https://aitmpl.com/trending/ Cleaning up the context window and letting Jev decide which Skill Claude Code should use seems to be gaining traction Have you tried it yet? How’s it working for you? Here’s the Mod: https://aitmpl.com/component/mod/productivity/jev-skill-suggestion

Opus 5.5 is a beast at making mods. It ported DK Spaceport from Mario Kart World into Mario Kart Wii, playable on the Mac via Dolphin emulator. It used Wiimm's modding tools to unpack the tracks and other object files from the game. Since Wiimm only ships Intel Mac builds, Opus compiled them from source with some linker tweaks to get the toolchain working on my Apple Silicon Mac. Opus then proceeded to build the entire track in Python and compiled the entire course data, collisions, checkpoints, CPU routes, and item boxes into the game's file format. It even referenced how the opening camera pan should work by decoding the files of other tracks in the game. For testing, Opus wrote a virtual controller and a bot that watches the game and feeds the buttons that need to be pressed during the gameplay. Dolphin exposes a debugger port for inspecting the emulated Wii's memory while the game is running, allowing the bot to understand the kart's game state at all times. As it's doing so, it will periodically take screenshots along the track that Claude will evaluate for improvements. This opens a whole new door for modding. You can use this to virtually create infinite game levels by reusing assets, in just tens of dollars and some hours of background work.

Play chess against Claude Code while you work I made this Mod that lets you play chess against Claude while your session keeps running It opens a sidebar with the board, and you can make moves by typing them or clicking with your mouse. Every Claude move shows how many tokens it used, with the full move history recorded. Install it with one command: npx claude-code-templates@latest --mod games/chess Everything is open source: https://github.com/davila7/claude-code-templates Enjoy the weekend building and playing chess with Claude!

aynısını claude code için mod olarak hazırladım (: tool-coach Model hata alan bir tool çağrısını aynı input ile tekrarlarsa mod çağrıyı durdurur. Bir dosya ya da bir komut bir şeyi değiştirene kadar çağrı yeniden çalışmaz. Model bunun yerine daha önce aldığı hatayı okur. kurulum için: https://cc-mods.keremgok.tr/mods/tool-coach.html

📑 GitHub - nwiizo/rust-learning-lab: Learn and review Rust through syntax, rationale, practice, and debugging. A bilingual plugin for Codex and Claude Code. 🅱 はてなブックマーク テクノロジー新着 より https://github.com/nwiizo/rust-learning-lab

Rustの学習についてです。 / 1件のコメント https://b.hatena.ne.jp/entry/s/github.com/nwiizo/rust-learning-lab#utm_campaign=bs_tw “GitHub - nwiizo/rust-learning-lab: Learn and review Rust through syntax, rationale, practice, and debugging. A bilingual plugin for Codex and Claude Code.” (3 users) https://htn.to/DYACWu7j5H

From JevJev to Jeff: our game-playing AI now runs on my own PC: For the past week I've been building an AI that plays Road to Vostok, a hardcore survival shooter. It started as JevJev. Its decision-making came from Jev, a model made by TypeSafe that picks the best option from a list and says how sure it is. Twice a second the mod asked Jev what to do: move, shoot, loot, run. A second, slower "overseer" asked Jev every 10 seconds which goal to pursue and whether it was stuck. It worked, kinda, and it was a lot of fun to watch. It also ran in the cloud, cost money every hour, and each answer took anywhere from half a second to four seconds. When I ran out of cloud credit, we switched. Jev's style of model has an open counterpart called Decider, and its smallest version runs on my own graphics card. The AI is now called Jeff, and his whole brain runs locally. There's no cloud bill, and most answers come back in a fraction of a second. Then we gave him a place to practice: a custom arena inside the game, 240 meters across, with enemies that keep coming, stocked military crates and airdrops, quest items, and all three of the game's traders in their own buildings. The Generalist's garage sits in the middle, and his attic above it is a working shelter. You walk up the ramp, go in, store your gear, and come back out into the arena with everything you carried. Jeff now completes the whole loop on his own. He searches crates, puts on the best armor he finds, walks into the shop, barters his loot for food and magazines, and hands in quest items. He isn't a smart shopper yet. He has overpaid, sold his own ammo, run away from fights he could have won, and at one point ate cat food he didn't need. But it's all him. Teaching Jeff to play like me: The newest part is that I play too, and Jeff learns from it. Every tick of my play is recorded, and we compare how I play with how he plays: - I loot crouched about 95% of the time. Jeff was at 4%. He now crouches on the way into a container. - I crouch about half the time when an enemy is in sight. Jeff never did. He now crouches when he stops to shoot. - I never eat or heal with a container open, because the game locks you into that screen until the timer runs out. Jeff spent over two minutes of one session doing exactly that. Now he treats himself only in his own inventory. My recorded play also goes straight into training. We use LoRA, a small add-on layer trained on top of the local brain. The base model stays untouched, and the add-on teaches it to answer Jeff's questions the way the recordings answered them. It trains on the same graphics card that runs the game, in about half an hour. If LoRA is new to you too, AI Coffee Break with Letitia's video "What is LoRA? Low-Rank Adaptation for finetuning LLMs EXPLAINED" is what made it click for me: http://youtube.com/watch?v=KEv-F5UkhxU The most useful lesson so far: scoring well against recorded answers doesn't mean playing well. One version agreed with my play more closely than any other, then did worst in the arena, because it ran from almost every fight. The arena, scored on survival, kills, loot and trades, is the real judge. The loop that improves Jeff from the outside: LoRA trains Jeff's brain. Everything around the brain improves through a second loop, and that one runs on Claude Code. It started on Fable 5.1 and moved to Opus 5.5 partway through. Round after round, Claude watches what Jeff actually did, measures it against my play and against the arena scores, finds why something went wrong, fixes the mod or the engine, restarts the game, and watches the next rounds. Some of what that loop caught: an enemy stuck inside a wall that could shoot out but couldn't be hit, a trader who was invisible because the game hides traders until a schedule opens them, a full backpack that made Jeff try to pick up the same helmet 374 times, and a Jeff who wasn't healing because he never opened his own inventory while hurt. Each one was found in the recordings, fixed, and checked in the next rounds. How Jeff is wired into the game: Jeff drives the game only through normal inputs: key presses, mouse movement and the game's own menus. Nothing is teleported or cheated. He only knows about enemies he can actually see or hear. F8 hands the character to Jeff, and any key or mouse movement from me takes it straight back. Some rules the brain can't talk its way past. He only fires when an enemy is actually in the gun's line of fire, and never at traders or the cat. He never drops gear he's wearing. Every trade is fair barter. A container screen closes the moment a fight starts. Even without the brain, the mod does real work. It records every tick of play, mine included. It shows an overlay of what the game sees and what Jeff decided. It runs the whole arena: the enemies, the starting kit, the traders, the stocked crates, the quests and the shelter. And if the brain is slow or offline, a simple rules-based player keeps Jeff moving. Where Jeff came from: Jeff didn't start from scratch. Two earlier projects of mine, both public on GitHub, got here first: - ai-plays-as-ai-making-paperclips (http://github.com/TechLuddite/ai-plays-as-ai-making-paperclips): a small AI running on my own PC that plays Frank Lantz's browser game Universal Paperclips by reading the page and clicking its buttons. No cloud, no API keys. It's the first time I had a local AI actually play a game but JS was used for the fast decisions. - Universal-AI (http://github.com/TechLuddite/Universal-AI): an idle game about optimization, and about who is doing it. You hand the wheel to an AI "Overseer", either a simple set of scoring rules or a small Llama model running right in your browser, and watch it drift from the goals you set. Both ideas carried straight into Jeff: he has an overseer that picks his goals, and a simple scoring fallback that keeps him playing if the AI brain is slow or offline. By the numbers: These come straight from Jeff's recordings and logs. A few are soft: the Jev count only covers sessions whose logs survived, the local count starts when we began logging every answer, and the arena rounds include some I played myself and a few short ones after restarts. - Jev in the cloud: about 12,200 decisions and 13 million tokens read, for about $3.50 in total. - Jeff on my PC: 131,500 decisions and 140 million tokens read in two days, for $0 in cloud fees. - Answer speed: 0.4 to 4.4 seconds with Jev in the cloud, and usually 0.1 to 0.3 seconds locally. - Arena rounds played: 276. - Gameplay recorded: 102 sessions, about 7 GB, turned into 227,000 practice examples. About 13,400 of those are my own choices of what to press and where to go. - Claude Code, running that outer loop: 19 working sessions since 6 September, 2.5 million tokens written, and 1.78 billion tokens read. Nearly all of that reading is Claude going back over its own working context at each step: my recording, Jeff's recordings, the project docs and code it wrote, rewrote and re-read many times over. Shout-outs - Road to Vostok, the hardcore single-player survival game Jeff lives in, made by Antti Leinonen: @roadtovostok on X, http://roadtovostok.com - Frank Lantz, whose Universal Paperclips my very first game-playing AI played: @flantz on X - TypeSafe, for Jev and the "System One" idea behind it: @typesafeai on X, http://github.com/typesafe-ai - Mark Marosi, for Decider, the open models Jeff runs on today: @notmapika on X, http://github.com/Mapika - AI Coffee Break with Letitia, for the video that explained LoRA to me: @AICoffeeBreak on YouTube - Omarchy, the Linux I run all of this on, by DHH: @dhh on X, http://omarchy.org, http://github.com/omacom/omarchy - Godot, the free and open-source engine Road to Vostok is built on, and the engine the phase 1 plug-in will be built for: @godotengine on X, http://godotengine.org What's next: This is phase 0, a small private proof of concept. Phase 1 is the real thing: an MIT-licensed plug-in for the Godot engine. It will ship with two things built together from scratch: a small decision model made just for games, and the engine wiring that lets it see, hear and act through a game's own controls, the way Jeff does today. The goal is for it to run on modest gaming hardware, not a workstation. What could that be? A player that plays alongside you, a companion that keeps you company in a survival game, or the greatest NPC nemesis of all time, one that learns how you play and comes back smarter every time you beat it. And why stop at one? Picture an army of Jeffs: whole squads, towns and rival factions, each one making its own decisions. In online games, it's for NPCs that the game's own developers choose to put there. The use I'm most excited about is accessibility. Picture voice-driven play: someone who can only use their voice tells Jeff where to go and what to do, and Jeff handles the keys and the mouse. With just their voice, they could explore so many wonderful worlds, in exactly their own way. "Look out! Jeff, RUN!" :D The same recipe could help my Opinionated-Omarchy project too. Its 505 audited Linux troubleshooting records are exactly the kind of clean data a LoRA needs, and its skill bench grades real outcomes the way the arena does for Jeff: its agent lane checks the machine afterwards. The prize is a small local model that fixes an Omarchy machine offline without breaking it. I've watched heavily shrunk local models run shell commands and break the very systems they were meant to fix. I'd love to see that solved on a machine with little memory, and certainly without the cloud or a monthly bill. That part is still an idea, not tested yet.

Cleaned up https://github.com/pathmodeio. Check an intent.md → CLI Draft specs in Claude Code → plugin Cursor/Codex/Desktop → MCP server Validate in CI → Action Implement the format → IntentSpec Start here: https://github.com/pathmodeio

clau"deco"de と言いたかっただけなのに mods どころか Ghostty ごと盛られてしまった私の claude code

マイクラの湧き潰し可視化MODが重くて使えないという人がいたので、もっと軽量なものをclaude codeで作ったんですが、ModrinthはAI生成のプラグインを公開するの禁止なのか 時代に逆行してるのでは CurseForgeで公開するか

Claude Opus 5.5は動画強いと言われているけど本当にそう思う。 このClaude Modsのデモビデオも「こんな感じで〜」みたいに指示したらほぼ一撃で作ってくれた

ModsはClaude Codeを使うなら知っておきたい機能。 Claude Mods 入門 | Claude Codeを自由にカスタマイズする｜nogu https://zenn.dev/nogu66/articles/claude-code-function-hooks-claude-mods #zenn