In this video, Chase AI argues that the most valuable part of an agentic operating system (AIOS) built on Claude Code is not the visual dashboard you can see, but everything happening under the hood: loop engineering, skill architecture, state management, and a coherent second brain. He breaks the AIOS construct into four levels and stresses that levels one and two hold roughly 90% of the value. Crucially, the skills you learn building an AIOS transfer to any project you tackle with Claude Code, or even other models like Codex or a local model.
When people see a polished AIOS, they split into two camps. The first is dazzled by the visual spectacle: buttons, moving parts, and metrics you can't find in the Claude Code terminal. The second dismisses it as smoke and mirrors that moves nothing forward. Chase says both camps miss the point: it is the AI fundamentals working under the hood that turn a fancy-looking web app into a customized weapon you can point at any problem.
Levels one and two are where "most of our money" is made. Once those are locked in, levels three and four are the "cherry on top."
Level one has four sub-phases: workflow audit, skill creation, automation, and loop engineering.
Before creating skills, you must know what to create skills for. Skills are powerful because they tell Claude to do a specific thing, in a specific way, for a specific output. Yet most people can't name the specific outputs they need repeatedly, and even fewer have turned those manual workflows into skills. Chase organizes his own work into domains (research, content, community, agency, sales) and notes that each domain contains repeated tasks, such as drafting outlines, finding video hooks, and repurposing content, that should be skills.
The mental model: treat Claude Code like a personal assistant you are onboarding. You would tell an assistant what you do and give them step-by-step instructions; that is exactly what codifying into skills accomplishes. Because skills are tangible, you can inspect and edit them until the outputs are dialed in.
If a skill gets repeated, turn it into an automation. You can prompt Claude Code directly ("Can we turn that skill into an automation?") or use a more visual approach in the Claude Desktop routines feature: name the routine, set the instruction to "run this skill," and put it on a schedule. On top of automations sits loop engineering, adding a self-improvement loop so a run can see how past iterations performed and improve future runs. This ties directly into level two, because the loop needs a place to record and read past runs.
Most of the discussion uses Obsidian as the example, but it doesn't have to be Obsidian; anything Obsidian does can be done in a traditional database. Obsidian is popular because it's free and easy to understand. The deeper point is file structure: if you set up Claude Code with a coherent, sensible file structure, you're "99% of the way there" even without a database or Obsidian.
Download Obsidian and designate a folder on your computer as the vault (new or existing). Whatever you store as the vault should contain at least a copy of the data for the domains you want help with (for example, all your sales data). To connect Claude Code, simply navigate your terminal into the vault folder and open Claude Code there; it is now connected to the vault.
The value of Obsidian is being able to ask Claude Code a question about any file across many subfolders and get a fast, accurate answer. If you dump millions of files into one folder with no hierarchy or backlinks, Claude struggles, which means it is slower, uses more tokens, and costs more money. Your job is to give Claude Code a clear map through the knowledge graph.
Chase references a tweet from Andrej Karpathy that earned over 20 million views, describing how to structure an Obsidian knowledge base so large language models like Claude can retrieve information quickly. The structure uses a primary vault with three subfolders:
The flow is: research lands in raw, becomes a wiki article in wiki, and turns into a deliverable in outputs (unstructured → structured → outputs).
The genuine power of the Karpathy system is an index.md file at every level. This markdown file acts as a table of contents that tells Claude Code what it's looking at as it descends each level. When you ask for information about a topic, Claude hits the top-level index.md, learns that wiki holds structured data, goes to the wiki folder, hits its index.md, and is routed to the right article. With thousands of documents, these index files make navigation dramatically faster and cheaper.
Chase stresses that the specific folder names are arbitrary. You don't have to use raw/wiki/outputs; you just need a map that makes sense for your data, which will be unique to you. If you're unsure, ask Claude Code to look at your vault and propose a structure, optionally using Karpathy's setup as inspiration. He also recommends a CLAUDE.md file in the vault that documents your vault conventions and a navigation pattern (the path Claude should follow to find things). His own vault has folders like content, notes, runs, inbox, ops, and projects.
Level two feeds loop engineering: outputs need to be logged somewhere Claude Code can later read, so self-improving skills and automations can see what past runs did and improve. Master levels one and two and you have 90% of an AIOS's power, all achievable through the terminal or desktop app, having codified your workflows and built a second brain.
Level three wraps everything from levels one and two in a custom visual layer. It can be web-app based or Obsidian based, with Claude Code still running under the hood connected to Obsidian. Chase's command center shows custom metrics tuned to his needs: YouTube subscriber count, Instagram, latest video, his Claude five-hour window, directives pulled from Google Calendar, and documents. Skills and automations become single buttons; clicking "inbox brief," for example, queues a headless Claude run that triages his inbox and creates drafts. He even attaches a local voice model (running on his own computer, free, not going out to ElevenLabs) so he can talk to it and have it talk back. In one demo it reports "32 threads triaged" with specific items flagged urgent.
The Obsidian version is similar, showing metrics like token burn with the same button-driven skill and automation runs and customizable tabs.
Building the web app is like building any web app with Claude Code: give it a visual reference (a screenshot of a layout you like), tell it the skills you use, ask it to connect to the vault, and specify the metrics you want in one place. For Obsidian, which runs a plugin system, you can ask Claude Code to turn the web app into an Obsidian plugin and install it.
Under the hood, clicking a button calls a headless version of Claude Code using the claude -p command, so the terminal never visibly pops up. Chase notes the brief drama where Anthropic said claude -p would draw from a $200 API credit rather than your Claude Max subscription, but they walked that back, so it currently still pulls from your Max plan, the same as running it in the terminal.
An AIOS doesn't have to be personal. You can share it with teammates or clients to raise the floor in your organization, since the power lives in the skills and automations behind one-click buttons. Recipients don't even need to run Claude Code themselves.
The customization is a major selling point for client work. Many people want to use AI and Claude but are scared off by the terminal and even the desktop app; for "99% of people" that's a bridge too far. Handing them a dashboard they can drive with a few buttons or their voice changes how non-technical users interpret these tools. Chase calls this dashboard effect something that "genuinely needs to be studied."
index.md maps at every level gives Claude a path that is faster and cheaper than a flat pile of files.