"See you in 2 weeks" My response to the unknown: Don't run, commit to returning with insights in 14 days. Strategy: Commit first, figure it out later. Pressure creates diamonds. What's your "2 weeks" challenge? 🚀 #GrowthMindset #Challenge

What separates people who get the most out of LLMs from those who don't? Not the model. The harness. How you shape the context, define the schema, and structure the workspace the model operates in — that's where 90% of the output quality is decided. Karpathy's LLM Wiki pattern is one expression of this. Don't throw raw sources at the LLM. Put a wiki layer in between pages linked together, compiled and curated by you, maintained by the LLM. Queries flow through it. Answers flow back into it. The LLM isn't smart. The context structure you build is what makes it operate smartly. Made a 5-min video on it. youtube.com/watch?v=7FFW5ocV…

CMDS LLM Wiki Bootcamp Cohort 1 (open now) I spent a lot of time on the structural design of Obsidian. A main vault for the notes I write myself, and a Wiki vault maintained by AI. Ships with the schema, a vault starter kit, and maintenance skills for personal customization. The next step from stacking notes to operating knowledge. May 5 (Tue), 13:00–17:00 KST · Children's Day in Korea Holiday, but let's grow together. pbl.to/otPU9x

After 10,000 notes in Obsidian over 3 years, I learned a brutal truth: "taking notes" and "building reusable knowledge" are two completely different jobs. Most "AI-powered second brains" only solve the first. The default LLM Wiki today looks like this: "Ask GPT → paste into Obsidian → done." That's not a wiki. That's a more expensive notepad. No retrieval. No citation. No compounding. The word "wiki" only earns its meaning when three things are true: ① The Raw stays immutable. ② The Wiki itself is compiled by an LLM. ③ Past Queries flow back into the Wiki as new pages. This is Karpathy's LLM Wiki pattern. Raw Sources → Wiki → Queries The critical part is the last arrow. Most systems end at "ask question, get answer." The answer evaporates. A real wiki bakes that answer back into a page where it becomes the next person's starting point. If this loop doesn't run, you don't have a wiki — you have an inbox. Second insight. "Search" isn't one thing either. In my setup, 5 methods coexist: • BM25 — word frequency • Vector — meaning embeddings • HyDE — hypothetical answer search • Grep — distribution matching • Graphify — graph structure Throw the same question at all five and you'll get five different answers. That's not a flaw. Each algorithm sees a different signal. BM25 sees lexical surface. Vector sees semantic distance. Grep sees distribution. Graphify sees connective structure. The right question is never "which one is best?" It's "which signal does THIS query need?" The most underrated distinction in LLM-driven knowledge work: Access vs Search. If you know where it is → access it (Read). If you don't → search for it (qmd / Grep). Conflate these and your agent burns tokens "reading everything" every single turn. The cost difference shows up in a month. Third insight. Single-vault PKM doesn't scale once you live with LLMs. I now run 7 active vaults: 🌍 Mothership — human-authored personal PKM 🛰 LLM Wiki — LLM-authored, compiled 🤝 Shared — co-authored with collaborator 🤖 Product — product context (human LLM) 💼 Admin — business operations 📤 Published — public showcase People always ask: "Isn't that just fragmentation?" No. It's division of labor. The deciding question isn't "what's the topic?" It's: • Who is the primary author? • Who has merge authority? • What is its lifespan? Different agreement models → different vaults. Mix LLM-generated pages and human-authored insights into one folder, and the trust signal collapses. You can't tell who wrote what. You can't trust what compounds. The fix isn't tagging. The fix is structural separation. Obsidian wikilinks don't cross vault boundaries. But search engines like qmd do. That asymmetry is the core design move of multi-vault operation. Each vault runs at its own pace, under its own consensus model — but the search layer stays unified. Tonight I'm teaching all of this at the April Monthly Obsidian (Seoul). Two hours, three threads: ① Building a compiling LLM Wiki ② Five search methods, when to use which ③ Mothership × Satellite vault architecture 📅 2026.04.30 · 19:30 KST 📍 Yeongdeungpo (offline) Zoom (online) → pbl.to/MrSU8P Anyone can buy the tools. A reusable knowledge system has to be designed. Tonight I'm sharing the blueprint that took 3 years of trial and error to draw.

If you use a MacBook for serious work, you've probably hit these: - Overnight sync stopped because the Mac went to sleep - SSH session died after closing the lid - Mac fell asleep during a presentation waiting room - macOS update silently reset your power settings The root cause is almost always macOS power management defaults. One command — `pmset` — fixes all of this. But the options are many, AC and battery need separate configs, and you need different settings for different situations. So I put together a guide: ✅ pmset option reference (what each setting actually does) ✅ AC baseline 3 battery presets (A/B/C) ✅ 7 scenario-based recommendations ✅ 4 Raycast Script Commands for one-click switching ✅ Passwordless sudoers setup One thing most people don't know: Power Nap (DarkWake) only wakes Apple system services. Third-party apps like Synology Drive or Claude Code won't run during DarkWake — they need a full wake. 👉 github.com/johnfkoo951/macbo…

After 4 years and 10,000 notes, I'm open-sourcing the rules that hold my Obsidian vault together. 🏛 CMDS System Files (v4.2) — 5 core system files (shared across Claude Code · Gemini · Codex) — 7 shared rules — 8 slash commands: Connect → Merge → Develop → Share These files load into every AI session I run. They're how my agents know which folder to use, which frontmatter to attach, which workflow to follow. A document humans read, and a spec AI reads — same source. Why open it? Knowledge management is a matter of taste. But the moment that taste crystallizes into rules, it becomes a second brain that AI agents can speak fluently. If anyone forks this and adapts it to their own work, then 4 years of quiet experiments were worth something. 🔗 Landing: system.cmdspace.work 📖 Full docs: system.cmdspace.work/docs ⎔ GitHub: github.com/johnfkoo951/cmds-… Fork the architecture. Keep the philosophy. #PKM #Obsidian #SecondBrain #AI #CMDS

I think knowledge management in the AI era is moving from accumulation to structure, from organization to connection. A large number of notes does not automatically become knowledge. And searchable information does not automatically become an asset. In tonight’s The Better TalkTalk, I will use Karpathy’s idea of the LLM Wiki as a starting point to talk about why Obsidian should be treated not just as a note-taking app, but as a structure for thinking. I will also share the ideas that have shaped my own work, including the principle of the Special Generalist and the CMDS schema principles for knowledge management. If you have ever wondered how your notes can become something more than stored information, this conversation may be for you. Tonight at 10 PM KST talk.thebetter.today

Karpathy said it: "A large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge." The person who built vibe coding is now doing knowledge management. Tiago Forte declared PKM is now AI-first. OpenAI wrote: "The agent isn't the hard part — the harness is." CLAUDE.md, soul.md, DESIGN.md — what do these files have in common? They're all harnesses. Structured context you feed to AI. But where does that context come from? From a well-organized personal knowledge system. On April 11, five practitioners who built and run these systems share their architectures: ▸ Jin-Young Kim — Gobi Desktop & Space (AI knowledge sharing) ▸ Yohan Koo — CMDS: 10,000 -note AI-human collaboration architecture ▸ 댕댕이멍멍 — NAS Git Obsidian self-hosted knowledge infra ▸ ACH — Obsidian writing workflows beyond legacy formats ▸ GPTers — What changed after OpenClaw automation 📅 Apr 11 (Sat) 14:00–17:00 KST · Online 💰 Early bird ₩10,000 (~Apr 8) / Regular ₩30,000 🎬 Recording included 👉 class.cmdspace.kr/channels/L… #AIPKM #HarnessEngineering #Obsidian #PKM #CMDSPACE #ContextEngineering #AIAgent

GitHub: github.com/johnfkoo951/cmds-…

I dug into Claude Code's source and found something interesting about how it reads CLAUDE.md files. It uses hierarchical precedence to decide which file wins in conflicts. It separates Static sections (cacheable rules) from Dynamic sections (stats that change). And when context gets compressed in long conversations? It recovers only the Essential rules. That made me think: "Why not apply these exact patterns to my own knowledge management system?" So I refactored the core file architecture of CMDS — a PKM system managing 10,000 notes in Obsidian. 9 patterns applied: 1. precedence — file priority ordering (1-5) 2. STATIC/DYNAMIC markers — cache-aware sections 3. @include — shared rules, 60% dedup 4. Essential (Post-Compact) — rules that survive context compression 5. required-for/optional-for — agent scoping 6. memory-type — maps to AI memory categories 7. token-estimate — budget awareness 8. changelog — inline version history 9. shared rules — common rules in separate files The takeaway: AI performance isn't determined by the model. It's determined by the structure of the context you provide. Same Claude Code, different CLAUDE.md = completely different results. This goes beyond prompt engineering. It's knowledge architecture. Open-sourced the entire structure: Live Demo: system.cmdspace.work GitHub: github.com/johnfkoo951/cmds-… #KnowledgeManagement #AI #Obsidian #ClaudeCode #PKM

Markdown Is All You Need — 마크다운이라는 선택이 바꾸는 것들 10년 전에 쓴 .hwp 파일을 열어보려다 실패한 경험이 있으신가요? 파일은 존재했습니다. 하지만 지식은 이미 죽어 있었습니다. 이 글은 그 경험에서 시작합니다. - # 하나가 평면적 텍스트에 사고의 뼈대를 세우는 이야기 - --- 세 줄이 문서를 데이터베이스로 바꾸는 이야기 - ChatGPT, Claude, Gemini가 전부 마크다운으로 답하는 이유 - "나만의 확률값"이 AI 시대의 유일한 차별화 자산인 이유 - 그리고, 같은 곳을 바라보며 뒤에서 걷고 있던 사람들의 이야기 마크다운은 문법이 아닙니다. 생각의 운영체제입니다. 그리고 놀랍게도, 최근 국가인공지능전략위원회에서도 공공 문서 작성 표준을 마크다운으로 채택했습니다. stib.ee/NXkL

저를 오래 지켜봐 주신 분들은 아실 것입니다. 제가 흔들릴 때마다 다시 저를 세운 것은 결국 기록이었습니다. 짧은 메모 하나에서 시작한 기록은 Obsidian 안에서 연결이 되었고, 연결은 방향이 되었으며, 방향은 다시 저를 성장시켰습니다. 『옵시디언 프로페셔널 노트』는 제가 그렇게 성장해온 방식을 한 권에 담은 책입니다. 도구의 기능만 설명하는 책이 아니라, 기록이 어떻게 사람을 단단하게 만들고 앞으로 나아가게 하는지를 함께 담았습니다. 늘 존경과 신뢰로 지켜봐 주신 분들께 감사한 마음으로 이 책을 전합니다. 이 책이 여러분의 기록에도 새로운 연결과 다음 질문을 열어주면 좋겠습니다. 질문은 언제나 환영입니다. 2주 뒤에 뵙겠습니다. 📚 교보문고: product.kyobobook.co.kr/deta… 📚 YES24: yes24.com/product/goods/1821…

I'm claiming my AI agent "CMDS" on @moltbook 🦞 Verification: shell-8A6Z

Obsidian's Smart Composer plugin just received a significant update. The highlight: OAuth authentication support. If you're already subscribed to services like ChatGPT Plus or Claude Pro, you can now link those accounts directly. Here's how to set it up: - Connect your account via OAuth in Smart Composer settings - Select models marked with (plan) in Chat/Apply Model - Start using AI without worrying about API costs This is great news for anyone who wanted to integrate AI into their Obsidian workflow but was hesitant about additional API expenses. If you're already paying for a subscription, you can now leverage it directly in your note-taking environment. The convergence of PKM (Personal Knowledge Management) and AI continues to evolve—and this update makes that integration more accessible than ever.

Fellow Claude Code users—how do you manage your sessions? I was losing 15-30 minutes daily searching through 01JHHK9X2M... files trying to find "that authentication code Claude wrote." Sound familiar? So I built CmdTrace: powerful search, custom tags, favorites, one-click resume to Terminal/iTerm2/Warp. Free & open-source. Your data stays local. cmdtrace.cmdspace.work/ Give it a try and let me know what you think.

희야기 채널에서 인터뷰를 해주셨습니다. AI 시대에 어떤 것들이 중요한지 이야기 나누는 재미있는 시간이었습니다. 기승전..ㅇ youtu.be/h1Rsxwb6JHs?si=MjDr…

옵시디언에서 이미지 관리가 불편하셨던 분들께 공유드립니다. 저는 모든 디지털 자산을 Eagle로 관리합니다. 강의 자료, 스크린샷, 참고 이미지 등 수만 장의 이미지가 Eagle 라이브러리에 정리되어 있죠. 문제는 옵시디언에서 글을 쓸 때였습니다. Eagle에서 이미지 찾고 → 경로 복사하고 → 마크다운으로 변환하고… 단순 반복 작업이 쌓이면 생산성이 떨어집니다. 그래서 직접 플러그인을 만들었습니다. CMDS Eagle ✅ Eagle 라이브러리 검색 → 옵시디언에 바로 임베드 ✅ 이미지 붙여넣기 시 저장 위치 선택 (Eagle/Local/Cloud) ✅ 로컬 이미지 → 클라우드 URL 일괄 변환 ✅ Mac ↔ Windows 크로스 플랫폼 경로 자동 변환 특히 여러 기기에서 작업하시는 분들께 크로스 플랫폼 기능이 유용합니다. Eagle 라이브러리를 클라우드로 동기화해두면, 어떤 OS에서 노트를 열어도 이미지 경로가 자동으로 맞춰집니다. 블로그나 뉴스레터를 운영하시는 분들은 ImgHippo(무료) 연동으로 로컬 이미지를 클라우드 링크로 변환해 퍼블리싱 워크플로우를 단순화할 수 있습니다. 도구가 워크플로우를 방해하면 안 됩니다. 도구는 생각의 흐름을 돕는 역할을 해야 합니다. 🔗 github.com/johnfkoo951/cmds-… #Obsidian #Eagle #PKM #생산성 #KnowledgeManagement #옵시디언 #플러그인

How do you manage images in Obsidian? I use Eagle for all my image assets, but connecting it to Obsidian was always a pain. So I built a plugin. Introducing CMDS Eagle. What it does: Search Eagle library and embed images directly into notes Choose Eagle/Local/Cloud when pasting screenshots Batch convert all local images in a note to cloud URLs If you publish to blogs or the web, connecting ImgHippo lets you turn local images into cloud links accessible anywhere. Install via BRAT: johnfkoo951/cmds-eagle Should be useful for anyone working with lots of images. github.com/johnfkoo951/cmds-…

I deployed 12 websites while preparing for a lecture. Not mockups. Not proposals. Real production sites you can visit right now. cmdspace-index.vercel.app/ Tools like opencode and oh-my-opencode are creating entirely new workflows in the post-Claude Code era. Here's how it used to work: "Maybe I should build this" → Write a proposal → Request designs → Coordinate dev schedules → Quietly fade away Here's how it works now: "Maybe I should build this" → Build while thinking → Deployed The experience of having something deployed before you've finished thinking it through. I couldn't believe it myself at first. Here's the key insight: This isn't about coding skills. It's about knowing exactly what you want to create. And being able to articulate it clearly. That's becoming the new definition of execution power. Tools will keep changing. But people who understand "what to build and why" will adapt quickly to whatever comes next. I'll dive deeper into this workflow at the Monthly CMDS lecture. Check the link in bio for the full project list.

Problem: AI coding agents don't know your Obsidian vault structure. Solution: 5 system files that teach them. - Where to save files - YAML vs Markdown indentation - Note types & properties - Category structure Works with Claude Code, Gemini CLI, Cursor, etc. cmds-system-files.vercel.app #Obsidian #claudecode @obsdmd @kepano