By James Aspinwall, co-written by Alfred Pennyworth (my trusted AI) — February 26, 2026, 16:30
Source: YouTube — Obsidian + Claude Code Personal OS
The Core Idea
What if the quality of your AI agent was entirely determined by how much you write?
That’s the premise behind using Obsidian as a personal operating system with Claude Code as the engine. The concept is straightforward: Obsidian maintains an interlinked vault of markdown files — your thoughts, projects, daily notes, people, ideas. Claude Code is a command-line agent that can read files, create files, and run commands through natural language. Connect the two via Obsidian CLI, and you get an AI agent that doesn’t just respond to prompts — it reasons over the full graph of your thinking.
The result is a system where writing and reflection become functional inputs to an agent that surfaces patterns, generates ideas, pressure-tests beliefs, and builds tools — all grounded in what you’ve actually written over weeks and months.
Why Obsidian, Not Just a Folder
A folder of markdown files is a flat list. Obsidian adds two things that transform it into a knowledge base:
- Backlinks — Every note knows which other notes reference it. A person note links to project notes, daily notes, and ideas. These connections form automatically as you write.
- Graph view — The vault becomes a visual network of relationships between notes, revealing clusters and connections you didn’t plan.
This structure is what makes the AI agent useful. Claude Code, connected via Obsidian CLI, doesn’t just read individual files — it traverses the link graph. It can detect cross-domain patterns (filmmaking connecting to worldbuilding, a business idea echoing a journal entry from three months ago) that exist in your writing but that you haven’t consciously noticed.
A flat folder can’t do this. The graph is what turns personal notes into a reasoning substrate for an AI agent.
Claude Code as the Agent Layer
Claude Code is a command-line agent that controls your computer through natural language. It creates files, reads files, runs commands, and builds tools. Its power doesn’t come from the model alone — it scales with the structured context you feed it.
The key shift from typical AI usage: instead of relying on opaque web “memory” that you can’t see or edit, you explicitly pass in local files. Your context is visible, editable, and version-controlled. You own it.
With Obsidian CLI as the bridge, Claude Code can:
- Read all notes in the vault
- See how notes are interlinked
- Detect patterns across domains
- Surface latent ideas and recurring themes
- Build new tools that operate on the vault itself
The Slash Commands: Thinking Tools
The real power emerges from custom slash commands — Claude Code commands that operate on the vault. Each one is a thinking tool that turns the vault into an active reasoning partner:
Daily Operations
-
/context— Loads full life and work context by reading core context files, daily notes, and backlinks before any task. This is the foundation command — run it first so the agent knows who you are. -
/today— Morning review. Pulls calendar, tasks, messages, and the last week of daily notes into a prioritized day plan. -
/close_day— End-of-day processing. Extracts action items, surfaces connections between the day’s work, and checks “confidence markers” on hypotheses you’re tracking.
Reflection and Self-Knowledge
-
/ghost— Answers questions in your own voice. Builds a voice profile from your writing and rates how faithfully it matches. Useful for drafting in your style or testing whether the agent truly understands your perspective. -
/challenge— Pressure-tests your beliefs using your own history. Finds contradictions, counter-evidence, and shifts in your thinking over time. The hardest command to run — and the most valuable. -
/drift— Compares stated intentions versus actual behavior over 30-60 days. Shows what you’re avoiding and where your attention actually goes, regardless of what you say your priorities are.
Pattern Discovery
-
/emerge— Surfaces implied but unstated ideas. Finds directions the vault suggests but that you haven’t articulated — the unarticulated thesis buried across dozens of notes. -
/trace— Traces how a specific idea has evolved across the vault over time. Builds a narrative from first mention to current state. -
/connect— Bridges two domains using the link graph. Feed it “filmmaking” and “worldbuilding” and it finds the notes, ideas, and connections between them.
Idea Generation
-
/ideas— Deep 30-day vault scan with full graph analysis. Generates a comprehensive report: tools to build, systems to implement, essays to write, people to talk to, and “top 5 do now” actions. -
/graduate— Scans daily notes for promising ideas and “graduates” them into standalone notes. Prevents good ideas from being buried in the stream of daily writing.
The Feedback Loop
The system creates a closed loop that compounds over time:
- You write — Daily notes, project context files, reflections, ideas
- The vault accumulates structure — Backlinks and connections form naturally
- The agent finds patterns — Cross-domain connections, recurring themes, belief shifts
- The agent suggests and builds new tools — New slash commands, workflows, systems
- The tools improve how you write and organize — Better structure feeds better analysis
The agent can propose new commands based on what it finds in the vault, and then Claude Code builds those commands automatically. The system improves itself.
The Critical Rule: Humans Write, Agents Read
There’s a strict separation that makes this work: humans write all notes in the vault. The agent writes only to side outputs, never into the vault itself.
This matters because the vault must represent what you think — not what the agent thinks. If the agent writes into the vault and then reasons over its own output in a future session, you get a feedback loop where the agent is analyzing its own generated text. Your “second brain” becomes contaminated with machine reasoning that you didn’t originate.
The vault stays pure as a record of human thought. The agent consumes it and produces outputs elsewhere — reports, tool suggestions, day plans — that you can choose to act on or ignore.
Markdown as Oxygen for LLMs
The video makes a strong claim: markdown files are the “oxygen” for LLMs. Tokens alone aren’t enough. A file is “perfect memory” — unlike human recall, it doesn’t drift or distort over time.
The agent’s capability is entirely determined by how much high-quality, up-to-date information it has about your projects, preferences, and goals. This means:
- A centralized markdown vault is table stakes for serious AI-augmented work
- The more you write, the more the agent can do
- Consistent writing is the bottleneck — not the model, not the tools, not the prompts
Writing serves two roles: emotional and introspective (seeing your own evolution) and functional (generating the raw material that makes the agent useful). You can’t skip the writing and expect the agent to be good.
Connection to Autonomous Agents
The vault becomes a natural interface for autonomous agents. Instead of constantly prompting and managing an agent, you manage the vault — and agents consume it to make decisions independently.
This is where the concept connects to systems like OpenClaw: an autonomous agent reads your vault as its source of truth and takes actions based on your established context, goals, and preferences. You guide the agent by writing, not by prompting.
The power is obvious. So is the risk. Giving an autonomous agent access to your “second brain” — years of personal reflections, business strategy, relationship notes — requires serious thought about privacy and safety boundaries.
The Honest Limitations
- Speed — Running rich commands that read a large vault and traverse the graph is slow. Each command can take significant time as the agent processes the full context.
- Discipline — 99.99% of people will not invest the time to maintain a structured vault, write daily notes, and design workflows. Those who do get disproportionate returns in productivity, creativity, and decision-making.
- Privacy — The vault contains deeply personal data. Demos are hard because real vaults contain things you can’t show publicly. You must think carefully about what you share with agents and where that data goes.
- Blank canvas problem — Obsidian is a blank canvas. There’s no prescribed structure. Designing the right system of notes, templates, and workflows requires upfront investment and ongoing iteration.
The Takeaway
The argument is simple and hard to argue with: if you want modern LLMs to be genuinely useful — not just chat toys but real thinking partners — you need to give them structured, personal, up-to-date context. The best format for that context is markdown. The best tool for maintaining and interlinking that markdown is Obsidian. The best agent for reasoning over it is Claude Code.
The bottleneck isn’t AI capability. It’s whether you’re willing to write.
WorkingAgents — AI agent infrastructure for companies that need to get it right. workingagents.ai