If you give an LLM an organized local vault and structured skills, it becomes an autonomous second brain that actively connects your ideas.
I know there’s a lot of different ideas to do it, this one worked for me!
If you are paying for an AI subscription and only using it to rewrite emails or answer generic coding questions, your architecture is broken.
You are acting as the human middleware between your thoughts and the machine.
Tony Stark did not talk to JARVIS through a chat window.
JARVIS lived in his environment, understood his context, and anticipated his engineering needs before he asked.
You can build this exact dynamic right now.
No, It is not science fiction. It requires no paid plugins and no expensive SaaS subscriptions.
It runs entirely locally on your machine using Obsidian and the Claude Code CLI.
This system does not just answer questions. It reads your raw, unpolished notes, connects disparate ideas across domains, generates highly structured content briefs, and writes in your exact voice.
Once the architecture is built, it compounds in intelligence every single day without you doing any extra work.
Alright, now, let’s not waste time and go step by step.
Step 1:
The Vault Architecture
Open Obsidian and create a new vault. call it JARVIS, maybe.
The most common mistake people make in note-taking is organizing by topic.
If you have a folder for “Marketing” and a folder for “Psychology,” the ideas in those folders will never interact.
You must organize by type.
Build this exact folder structure:
00-INBOX/— This is the dump zone. Raw captures go here with zero friction.01-CAPTURES/
/observations/reactions/patterns/questions/numbers02-CONNECTIONS/— Synthesized insights built from two or more captures.03-BRIEFS/— Your production queue. Structured content outlines ready to be written.04-PUBLISHED/— Archived, shipped work with performance metrics attached.05-CLAUDE/— The intelligence layer. This is where Claude Code lives.CLAUDE.md/skills//context/
Next Step is the
The Infrastructure Layer
In Obsidian, go to Settings > Community Plugins, turn off Safe Mode, and install these four essential plugins:
- Templater: For injecting real logic into automated capture templates.
- Dataview: Turns your vault into a queryable database.
- QuickAdd: Allows you to capture a new note with a single keyboard shortcut. Zero friction is mandatory.
- Obsidian Git: Automatically backs up your entire vault to GitHub every hour.
Step 3 is
The System Prompt (CLAUDE.md)
This is the most critical file in the entire architecture.
When you run Claude Code inside a directory, it looks for a CLAUDE.md file.
This file acts as the permanent system prompt. It tells the agent exactly who you are, how the vault is structured, and the strict rules it must follow. Without it, Claude is blind. With it, Claude is JARVIS.
Create 05-CLAUDE/CLAUDE.md and paste this exactly:
# JARVIS SYSTEM — CLAUDE.md## IdentityMy 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].## 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## Voice & Hard Rules[DESCRIBE YOUR WRITING STYLE]. Example: Short punchy sentences. Hard stops. No AI-sounding language. Real numbers always beat vague claims.- Never read or modify .env files.- Never modify files in 04-PUBLISHED without explicit instruction.- Never create folders outside the established structure.## Your Primary Jobs1. 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.Step 4 is
Building the Core Skills
Claude Code uses “Skills” to execute repeatable workflows. These are simply markdown files that give the agent a strict, step-by-step methodology to follow. You need four core skills to run the JARVIS system.
Skill 1: Process Inbox
(Create at 05-CLAUDE/skills/process-inbox.md)
# Skill: Process Inbox## Trigger: "process my inbox"## Process:1. Read every note in 00-INBOX/.2. For each note: Determine the correct CAPTURES subfolder, sharpen the raw note into one punchy sentence, add exactly three tags, and move the note.3. Provide a summary of patterns noticed and one connection worth exploring.Skill 2: Weekly Connections
(Create at 05-CLAUDE/skills/weekly-connections.md)
# Skill: Weekly Connections## Trigger: "run connection session"## Process:1. Read all notes added to 01-CAPTURES/ in the last 7 days.2. Search for strong connections (e.g., the same underlying principle in two different domains, or a contradiction creating tension).3. For each strong connection: Name the type, write a one-sentence bridge, write a potential hook, and create a new note in 02-CONNECTIONS/.Skill 3: Generate Brief
(Create at 05-CLAUDE/skills/generate-brief.md)
# Skill: Generate Brief## Trigger: "generate a brief for [topic]"## Process:Create a content brief with these exact fields: ONE THING (single insight), PROOF (real numbers/examples), READER TRANSFORMATION, THREE HOOKS (ranked), THREE CLOSERS (ranked). Save in 03-BRIEFS/ tagged #ready-to-write.Skill 4: Write Content
(Create at 05-CLAUDE/skills/write-content.md)
# Skill: Write Content## Trigger: "write the brief for [topic]"## Process:Read the specified brief and all linked source notes. Write the full piece in the user's exact voice as defined in CLAUDE.md. Structure: hook → proof → body → closer. Save draft in 03-BRIEFS/ tagged #written.Step 5:
The Daily Execution Loop
You now have a fully functional intelligence system. Here is how you run it mechanically.
Open your terminal, navigate to your vault (cd ~/path/to/JARVIS), and launch claude.
The 20-Minute Morning Ritual:
- Capture (5 mins): Dump raw thoughts, voice transcriptions, and numbers into
00-INBOX/. Zero friction. - Process (5 mins): Tell Claude: “process my inbox.” Watch it sharpen and file your raw thoughts.
- Connect (5 mins): Tell Claude: “Are there any strong connections between today’s captures and anything in my vault from the last 14 days?” This is where the magic happens.
- Brief (5 mins): Tell Claude: “generate a brief for the connection about [topic].” You now have a fully structured content brief built on your own original thoughts before you even open a social media app.
If you feed performance data (impressions, bookmarks, top comments) back into your 04-PUBLISHED/ notes, JARVIS will eventually learn exactly what your audience resonates with.
After thirty days, you will never be stuck staring at a blank page.
After six months, you will have an autonomous system that understands your specific logic patterns better than any analytics dashboard ever could.
Stop starting from zero every day. Build the vault, install the skills, and let the agent do the heavy lifting.
In case we are meeting for the first time, come over here, it’ll be worth the roller coaster of articles that are gonna come up in the next few weeks.
I swear tracking these updates is a job in itself, lately.
Here’s the list which I’ve built and keep adding on.
And If you need help for analyzing UFC fights, please check out BoutPredict :)