A practical guide to turning Obsidian into an AI-powered second brain using Claude Code. Store everything you know about your business in markdown files — clients, sales scripts, offers, meeting notes, voice profiles — and let a Claude agent read and write them directly from your terminal.
Full content
# Build an AI Second Brain with Claude Code + Obsidian
## What This Is
Your second brain is a folder of plain markdown files that a Claude Code agent can read and write from the terminal. Instead of scattered notes, CRM spreadsheets, and Google Docs you never find, everything lives in one vault — and a Claude agent acts as the intelligent interface on top of it.
This method was popularized by Israeli consultant Yuval Kashtecher, who wired up his entire business to this system, connecting it to email, meeting transcription (Spinach), a newsletter tool, and a payment processor.
---
## Why It Works
- **Markdown is universal.** Claude Code reads plain text natively. No API integrations, no database setup.
- **Claude Code runs locally.** You `cd` into your vault folder, type `claude`, and the agent has full read/write access to every file.
- **Context compounds.** The more you document, the smarter the agent gets. Ask "who hasn't paid?" and it reads `clients.md`. Ask "write a cold outreach email" and it reads `sales-scripts.md` + `voices/linkedin.md`.
---
## Recommended Vault Structure
```
your-vault/
├── clients.md # Every client: story, last contact, what they paid
├── sales-scripts.md # Call scripts, objection handling, closes
├── offers.md # Products, pricing, what works / what doesn't
├── meetings/ # One file per day: decisions, actions, context
│ └── 2026-05-12.md
└── voices/ # How you sound on each channel
├── facebook.md
├── linkedin.md
└── newsletter.md
```
---
## Step-by-Step Setup
### Step 1 — Install Obsidian
Download Obsidian (free) from https://obsidian.md and create a new vault folder on your computer. This folder is just a regular directory — Obsidian is the visual editor, but Claude Code works directly with the files.
### Step 2 — Install Claude Code
From your terminal, run:
```bash
npm install -g @anthropic-ai/claude-code
```
Then navigate into your vault and start Claude:
```bash
cd ~/your-vault
claude
```
The agent now sees every file in the folder. It can read, write, search, and summarize.
### Step 3 — Create Your First File
Start with `clients.md`. List every client, their story, what you discussed, what they paid. Then ask:
> "Read clients.md and tell me who's most worth my time this week."
From there, add files one at a time. The vault grows with you.
---
## Example Prompts Once Running
| Goal | What to type |
|------|-------------|
| Chase inactive clients | "Who in clients.md hasn't been contacted in 30+ days?" |
| Write a sales script | "Read sales-scripts.md + voices/linkedin.md and write a script for a new female client" |
| Prep for a meeting | "Read meetings/ and summarize open action items" |
| Price a new offer | "Read offers.md and suggest a price point for a new 3-month package" |
| Post content | "Read voices/facebook.md and write a post about AI second brains" |
---
## Connecting External Tools
Once the local vault is working, you can extend it:
- **Email** — Claude Code can draft and send via Gmail MCP
- **Spinach** — Auto-transcribes every meeting and produces a call script; feed the output into `meetings/`
- **Newsletter tools** — Export drafts from Claude into your mailing platform
- **Payment processors** — Log payment events into `clients.md` automatically
---
## PARA Method Integration
For a more structured vault, adopt Tiago Forte's PARA framework:
- **P**rojects — active work with deadlines
- **A**reas — ongoing responsibilities (clients, finance, health)
- **R**esources — reference material (scripts, offers, research)
- **A**rchives — completed or inactive items
Claude Code navigates PARA naturally — just describe what you need and it finds the right files.
---
## References
- [How To Geek: Claude + Obsidian Second Brain](https://www.howtogeek.com/claude-obsidian-the-cheat-code-for-building-a-second-brain/)
- [MindStudio: Build AI Second Brain](https://www.mindstudio.ai/blog/build-ai-second-brain-claude-code-obsidian)
- [PARA Method + Claude Code + Obsidian](https://aimaker.substack.com/p/para-method-tiago-forte-claude-code-obsidian-ai-productivity-os)
- [Claude Code Quickstart](https://code.claude.com/docs/en/quickstart)
- Original post by Yuval Kashtecher (WhatsApp AI community)
---
## Level Up: Clone Yourself with AI
*From the AI Cloning Guide 2026*
Once your second brain is storing your voice profiles and institutional knowledge, the natural next step is deploying an AI clone of yourself — an agent that sounds like you, knows what you know, and can act on your behalf across platforms.
### Step 1: Extract Your Style DNA via the "Taste Interview"
Inside Claude, run a structured "Taste Interview." Ask Claude to interview you about your communication preferences, storytelling style, sense of humor, and values. The output becomes your `voices/style-dna.md` — the master reference for all clone deployments.
### Step 2: Voice Cloning via ElevenLabs
Record 5–10 minutes of clean, expressive speech. Upload to ElevenLabs and generate a voice model. Store the voice ID in `voices/elevenlabs.md` so any agent can reference it when producing spoken content.
### Step 3: Deploy on 4 Platforms
| Platform | Best For | Setup | Key Feature |
|----------|----------|-------|-------------|
| **OpenClaw** | WhatsApp, multi-channel | Railway one-click or Docker + Composio | 200+ service integrations |
| **Hermes** | Self-improving learning clone | One-line install | SOUL.md persistent memory, multi-provider (Claude/GPT/Gemini) |
| **NanoClaw** | Security-sensitive data | 3-command install | Docker + micro-VM isolation, 700-line auditable codebase |
| **Claude Managed Agents** | Enterprise / API scale | Anthropic API | Structured agent loop, production-ready |
**Recommended stack:** OpenClaw for WhatsApp presence → Hermes for continuous learning → NanoClaw when handling sensitive client data → Claude Managed Agents for scale.
### Phase 4: Ongoing Maintenance
- **Feedback loop** — review agent outputs weekly; correct and feed corrections back in
- **/learn command** — teach Hermes new facts on the fly without rewriting the system prompt
- **/compress** — compress accumulated memory to keep the context window efficient
---
## Deploying Real Agents: Barry Zhang's 5-Step Methodology
*From the AI Agent Deployment Guide 2026 — Barry Zhang, Anthropic*
Barry Zhang (Anthropic) distilled production agentic deployment into a 5-step framework that maps naturally onto the second brain workflow.
### Step 1 — Evaluate the Use Case
Before building, weigh four factors: **complexity** (does the task have enough steps to justify an agent?), **value** (what's the upside if it works?), **viability** (can current models handle it reliably?), and **error cost** (what happens when it fails?). Only build an agent when complexity × value exceeds the error risk.
### Step 2 — Define Components
Every agent needs three things: an **environment** (what can it see — your vault folder, email inbox, calendar), **tools** (what can it call — file read/write, web search, email send, payment lookup), and a **system prompt** (who is it, what are its rules, what's its north star). In your second brain, this maps directly: environment = vault, tools = file ops + MCPs, system prompt = `CLAUDE.md`.
### Step 3 — Implement the Agentic Loop
The core loop: **receive task → read context → plan → execute → evaluate result → loop or stop.** Claude Code handles this natively when run inside your vault. For external deployments, OpenClaw and Hermes package this loop for you.
### Step 4 — Optimize via Agent's Perspective
Three techniques: **context review** (are the right files in scope? is the system prompt unambiguous?), **self-debugging** (ask the agent to explain its reasoning *before* it acts), and **trajectory analysis** (review past sessions to find where the agent drifted or made wrong assumptions).
### Step 5 — Deployment & Maintenance
Set token and cost budgets per session to avoid runaway spend. Evolve tools as your vault grows (add MCPs for email, calendar, CRM, payments). Scale with **multi-agent** setups: one specialized agent per domain (clients, content, finance).
### Platform Deep Dives
**OpenClaw** — Railway one-click deploy or Docker. Connect to 200+ services via Composio. Best entry point for WhatsApp automation and multi-channel presence. Start here if your goal is reaching clients where they already are.
**Hermes** — One-line install. SOUL.md acts as the agent's persistent memory (like your vault's `CLAUDE.md`). Supports Claude, GPT-4, and Gemini — switch providers without rewiring. Best for a clone that genuinely learns and improves over time.
**NanoClaw** — 3-command install. Docker sandbox with micro-VM isolation. At 700 lines of code it's small enough to audit yourself. Best when processing sensitive client data or financial records — you can verify exactly what it does.
**Claude Managed Agents** — Anthropic's API-native approach. Full control over the agent loop, tools, and system prompt. Best for enterprise workflows or when you need SLA-backed reliability at scale.
---
## Putting It All Together: From Second Brain to AI Clone
```
Your Obsidian Vault
├── CLAUDE.md ← Agent identity + rules (system prompt)
├── clients.md ← Knowledge base
├── voices/
│ ├── style-dna.md ← Taste Interview output
│ └── elevenlabs.md ← Voice clone ID
└── meetings/ ← Continuous learning feed
↓ Claude Code (local) ← Read/write your files
↓ OpenClaw (WhatsApp) ← Reach clients
↓ Hermes (self-improving) ← Learn continuously
↓ NanoClaw (secure) ← Handle sensitive data
↓ Claude Managed Agents ← Scale to enterprise
```
The second brain is the foundation. The agent platforms are the deployment layer. The voice clone is the interface. Together they form a complete AI operating system for your work.