Claude Code×ObsidianでJARVIS構築

Tony Stark did not build JARVIS in a day. But he also did not have Claude Code. What I am about to show you is the closest thing to a real JARVIS that exists right now. Not a chatbot you talk to. Not a notes app you type into. A system that reads your thinking, connects your ideas, surfaces what matters, and writes alongside you in real time. It runs inside Obsidian. It is powered by Claude Code. And once it is set up it gets smarter every single day without you doing anything extra. This is the full setup. Every step. Every file. Every prompt. Copy it exactly. What JARVIS Actually Is Most people think JARVIS is a voice assistant that answers questions. That is the surface. The real JARVIS is a system that knows its owner deeply enough to anticipate what they need before they ask for it. The version you are building today does four things. It captures everything you think, notice, and observe with zero friction. It connects ideas across your vault automatically, finding relationships you would never find manually. It generates content briefs from those connections so you never start from blank. And it writes in your exact voice from those briefs so the output sounds like you, not like AI. That is the system. Let's build it. What You Need Before You Start Obsidian — download at obsidian.md. Free for personal use. Your notes are plain markdown files on your machine. Claude Code — install with: npm install -g @anthropic/claude-code An Anthropic API key — get one at console.anthropic.com. Set your API key in the terminal: export ANTHROPIC_API_KEY=your_key_here That is everything. No paid plugins. No subscriptions beyond your Anthropic API usage. Step One: Build the Vault Architecture Open Obsidian. Create a new vault. Name it whatever you want. I name mine JARVIS. Now create this exact folder structure: JARVIS/ ├── 00-INBOX/ ├── 01-CAPTURES/ │ ├── observations/ │ ├── reactions/ │ ├── patterns/ │ ├── questions/ │ └── numbers/ ├── 02-CONNECTIONS/ ├── 03-BRIEFS/ ├── 04-PUBLISHED/ └── 05-CLAUDE/ ├── CLAUDE.md ├── skills/ └── context/ This structure is not arbitrary. Every folder has a specific job. INBOX is where raw captures land with zero organization. The goal at capture time is speed not structure. You sort it later. CAPTURES is organized by note type not topic. This is the most important architectural decision in the whole system. When you organize by topic a note about AI content strategy and a note about how attention works psychologically never meet. When you organize by type they both land in the patterns folder and Claude Code finds the connection between them automatically. CONNECTIONS is where synthesized insights live. These are not raw captures. They are new ideas that emerge from the relationship between two or more captured notes. This is where your best content comes from. BRIEFS is your production queue. Every file in here is a piece of content ready to write. Hook written. Closer written. Source notes linked. The actual writing is the last step, not the first. PUBLISHED archives everything you have shipped with performance data attached. This is how JARVIS learns what works over time. CLAUDE is the intelligence layer. This is where Claude Code lives when it is working inside your vault. Step Two: Install the Essential Plugins In Obsidian go to Settings, Community Plugins, turn off Safe Mode, then Browse. Install these four: Templater — runs dynamic templates with real logic inside them. You will use this for automated capture templates. Dataview — queries your vault like a database. Surfaces notes by tag, date, link count, or any property you define. QuickAdd — captures a new note with one keyboard shortcut. Zero friction is non-negotiable at capture time. Obsidian Git — backs up your vault to GitHub automatically every hour. Set it and forget it. Step Three: Build the CLAUDE.md File This is the most important file in your entire system. The CLAUDE.md file tells Claude Code everything it needs to know about who you are, what your vault contains, how it is organized, and exactly what you want it to do when it works inside your vault. Without a CLAUDE.md Claude Code treats your vault like any random folder. With a strong CLAUDE.md it treats your vault like a system it knows deeply and has been managing for months. Create this file at 05-CLAUDE/CLAUDE.md: # JARVIS SYSTEM — CLAUDE.md ## Identity My name is [YOUR NAME]. I am [WHAT YOU DO]. My audience is [WHO FOLLOWS YOU]. My content pillars are [LIST YOUR 2-3 MAIN TOPICS]. ## This Vault This is my second brain and content production system. Every note is raw material for content, thinking, or decision-making. Nothing here is decoration. ## Vault Structure - 00-INBOX: unprocessed captures — always check here first - 01-CAPTURES/observations: things I noticed, raw and unpolished - 01-CAPTURES/reactions: my honest gut response to something I read or heard - 01-CAPTURES/patterns: the same principle appearing in two different domains - 01-CAPTURES/questions: things I genuinely do not know the answer to - 01-CAPTURES/numbers: real data points with specific numbers attached - 02-CONNECTIONS: synthesized insights from two or more captured notes - 03-BRIEFS: content ready to write, structured with hook and closer - 04-PUBLISHED: archived content with impression and bookmark data - 05-CLAUDE: your working directory ## My Voice [DESCRIBE YOUR WRITING STYLE IN SPECIFIC TERMS] Example: Short punchy sentences. Hard stops. Every idea gets its own line. No dashes. No AI-sounding language. Real numbers always beat vague claims. ## Hard Rules - Never read, access, or modify any .env files - Never modify files in 04-PUBLISHED without explicit instruction - Never create folders outside the established structure - Never use [WORDS YOU HATE] in any content you write for me ## Your Primary Jobs 1. Process INBOX — read raw captures, sharpen them, file them correctly 2. Run connection sessions — find non-obvious relationships between recent captures 3. Generate briefs — turn strong connections into structured content briefs 4. Write content — produce finished posts in my exact voice from approved briefs 5. Log performance — update PUBLISHED notes when I provide data The more specific you make this file the better Claude Code performs inside your vault. Vague CLAUDE.md files produce vague outputs. Treat this as the most important document you write today. Step Four: Connect Claude Code to Your Vault Open your terminal. Navigate to your Obsidian vault: cd ~/path/to/your/JARVIS-vault Launch Claude Code: claude Claude Code will automatically find your CLAUDE.md file and load everything in it as permanent context. Every conversation you have inside this directory is now informed by your entire vault structure. Test the connection immediately: Read my CLAUDE.md and tell me in one paragraph what you understand about this vault and what your role is inside it. If the summary is accurate and specific your connection is working. If it is vague your CLAUDE.md needs more detail. Step Five: Build the Core Skills A skill in Claude Code is a reusable workflow stored as a markdown file. You call it by name and Claude Code executes the full process every time. You are going to build four skills. These four cover 90% of everything JARVIS does. Skill One: Process Inbox Create 05-CLAUDE/skills/process-inbox.md: # Skill: Process Inbox ## Trigger phrases "process my inbox" / "run morning capture" / "clear the inbox" ## Process 1. Read every note in 00-INBOX/ 2. For each note: a. Determine which CAPTURES subfolder it belongs to b. Sharpen the raw note into one punchy sentence c. Add exactly three tags — no more, no fewer d. Move the sharpened note to the correct subfolder 3. After processing all notes, provide: - Total notes processed and where each went - Any patterns noticed across today's captures - One connection worth exploring from today's batch ## Quality bar A sharpened note should be specific enough that a stranger would understand exactly what was observed without any additional context. If it still needs explanation it is not sharp enough. Rewrite it. Skill Two: Weekly Connections Create 05-CLAUDE/skills/weekly-connections.md: # Skill: Weekly Connections ## Trigger phrases "run connection session" / "find this week's connections" / "weekly connections" ## Process 1. Read all notes added to 01-CAPTURES/ in the last 7 days 2. Search for connections across ALL subfolders simultaneously 3. A strong connection is one of these four types: TYPE A: Same underlying principle in two different domains TYPE B: Contradiction between two notes that creates interesting tension TYPE C: Pattern connecting three or more notes into one unnamed insight TYPE D: A question from one note that another note accidentally answers 4. For each strong connection: a. Name the connection type b. Write a one-sentence bridge between the ideas c. Write a potential content hook using this connection d. Create a new note in 02-CONNECTIONS/ linking the source notes ## Quality bar If the connection is obvious it does not qualify. Only surface connections that would genuinely surprise the person who wrote the notes. Minimum: 3 connections. Maximum: 5. Quality over quantity. Skill Three: Generate Brief Create 05-CLAUDE/skills/generate-brief.md: # Skill: Generate Brief ## Trigger phrases "generate a brief for [topic]" / "brief this connection" / "make a brief" ## Process Create a content brief with exactly these five fields: ONE THING The single insight this piece is built around. Must be one sentence. If it cannot be one sentence the idea is not ready. Push back if the one thing is fuzzy. PROOF The most specific real example, number, or result that proves the one thing. Vague proof invalidates the brief. Real numbers only. READER TRANSFORMATION What does the reader know or feel at the end that they did not before? If this cannot be stated clearly the piece has no reason to exist. THREE HOOKS (ranked) Different in approach and tone. Rank by how hard they stop the scroll. Hook one should be aggressive. Hook two should be curious. Hook three should be personal. THREE CLOSERS (ranked) Rank by urgency and memorability. The closer is written before the middle. Always. ## Output Save as a new note in 03-BRIEFS/ with today's date and topic as filename. Tag it #ready-to-write Skill Four: Write Content Create 05-CLAUDE/skills/write-content.md: # Skill: Write Content ## Trigger phrases "write the brief for [topic]" / "write this brief" / "produce content from brief" ## Process 1. Read the specified brief from 03-BRIEFS/ 2. Read all source notes linked in the brief 3. Write the full piece in the user's exact voice as defined in CLAUDE.md 4. Structure: hook → proof → body → closer 5. Every section must add specific value. No filler. 6. Include at least two copyable code blocks or prompt examples where relevant 7. End with bookmark CTA and follow instruction ## Voice requirements (from CLAUDE.md) Apply every voice instruction precisely. If in doubt, make it shorter and more direct. The output should be indistinguishable from content the user wrote themselves. ## Output Save draft in 03-BRIEFS/ next to the source brief. Tag it #written Step Six: The Daily JARVIS Ritual This is how you run the system every morning. Total time: twenty minutes. Open your terminal. Navigate to your vault. Launch Claude Code. Minute one to five — capture: Before you open anything else, spend five minutes adding raw captures to 00-INBOX. Voice memo transcriptions. Things you noticed during your commute. A number you saw. A reaction you had to something you read last night. Raw. Unpolished. Just get it in. Minute six to ten — process: Run the process inbox skill. Read the report. Note what got filed and what patterns Claude Code noticed across today's captures. Minute eleven to fifteen — connect: Are there any strong connections between today's captures and anything in my vault from the last 14 days? This is the JARVIS moment. You are asking the system to think across your entire recent history and surface what belongs together. Minute sixteen to twenty — brief: Generate a brief for the connection about [whichever one surprised you most]. Done. You have a content brief ready before you open any social media platform. The rest of the day is execution not ideation. Step Seven: The Weekly Connection Session Every Sunday run this full session: Run the weekly connections skill on all captures from the last 7 days. Let Claude Code read the full week of captures across all subfolders simultaneously. The connections it surfaces from a full week of inputs will be stronger than anything you find in the daily session. Review the connection notes it creates in 02-CONNECTIONS. Pick the two strongest. Run the brief skill on both. You now have two fully briefed content pieces ready to write before Monday morning. Step Eight: Close the Loop With Performance Data This is the step that makes JARVIS compound over time. Every time a piece of content performs well, open the corresponding note in 04-PUBLISHED and add the numbers: impressions: 443,000 bookmarks: 11,678 engagement_rate: 0.04% top_comment: [paste the comment that sparked the most replies] what_worked: Netflix format hook + Obsidian + Claude Code topic combination Once a month run this session: Read everything in my 04-PUBLISHED folder. Tell me: 1. Which topics drove the most bookmarks per impression 2. Which hook formats outperformed their topic average 3. What three content angles my best performing posts suggest I have not tried yet 4. Which combinations of topic and format I should double down on this month Only give me insights specific to this exact data. No generic advice. This monthly session is what transforms JARVIS from a productivity tool into a content intelligence system. After six months it will know your audience better than any analytics dashboard because it is reading the content and the data together, not just the numbers. What This System Produces After thirty days you will notice you are never stuck for ideas. After ninety days you will notice your content is more specific, more connected, and more original than anything you were producing before. After six months you will have a vault that functions as a genuine second brain. One that has captured hundreds of observations, connected them in ways you never would have found manually, and produced content from those connections that your audience responds to because it reflects genuinely original thinking rather than recycled takes. That is the real product of this system. Not the notes. Not the productivity. The compounding intelligence that emerges from consistent capture, consistent connection, and consistent feedback over time. Most people are starting from zero every day. JARVIS makes sure you never do. Bookmark this guide and start the setup today. Follow @cyrilXBT for daily Claude Code builds, AI tools, and the exact systems behind everything I ship.