Quick Overview
This video is a build-along tutorial presented by Nate Herk, founder of Uppit AI and the AI Automation Society. Herk demonstrates how to build a personalized AI Operating System and second brain using GPT-6 Astra inside the Codex environment. The guide follows on from AI productivity frameworks to show viewers how to organize local knowledge files and autonomous agent workflows.
Key Points
- 1.An AI Operating System structures personal and business knowledge into a local repository that AI models can reference and manipulate directly.
- 2.The 4Cs architecture divides the system into Context and Connections for the second brain, alongside Capabilities and Cadence for the operating system.
- 3.The core operating rules and file routing maps reside in an AGENTS.md or CLAUDE.md file that the model reads before processing user prompts.
- 4.Automated skills including onboarding, audit, level up, grill me, and 3D visualization streamline knowledge collection and system maintenance.
- 5.Adopting Karpathy-style LLM wikis allows the system to organize raw inputs into cross-referenced Markdown files rather than relying solely on standard retrieval.
Summary
Nate Herk presents a walkthrough of how he built a personal AI Operating System and second brain named Herk Brain, powered by GPT-6 Astra inside the Codex desktop interface. He begins by demonstrating an interactive 3D universe visualization of his company's entire knowledge base, containing interconnected nodes for business wikis, meeting records, YouTube video knowledge, and agent skills. Because this knowledge is indexed and stored locally in files, GPT-6 Astra can act as an informed business partner with comprehensive memory across operations rather than a basic conversational chatbot.
To explain how the system functions, Herk introduces the 4Cs framework: Context, Connections, Capabilities, and Cadence. Context comprises stable, slow-changing business and personal data such as company structure, quarterly goals, and customer avatars. Connections represent dynamic daily tools like Gmail, Slack, ClickUp, and Stripe. These two components establish the second brain. Capabilities cover executable tools and skills, while Cadence introduces autonomous schedules and recurring routines. Herk emphasizes that adding capabilities and automated cadences without first establishing deep context leads to generic outputs, making rich documentation essential.
Herk walks through initializing an AI Operating System from scratch. In Codex, he creates a local project folder and runs an onboarding skill from a resource pack repository. An onboarding agent asks seven core questions regarding the user's role, business offering, voice, and goals, immediately creating dedicated Markdown files like about-me.md, about-business.md, and priorities.md in a context directory. It also creates an AGENTS.md file, which serves as the master instruction file and routing directory for Codex, functionally identical to CLAUDE.md used in Claude Code. The routing map specifies exact file paths for business wiki topics, voice guidelines, and project assets so the agent navigates directly to sources without guessing.
To expand and refine the repository, Herk demonstrates several packaged skills. Running the audit skill inspects the entire directory against the 4Cs framework, outputting an evaluation score and saving a timestamped report in an audits folder. The level-up skill reads this report to identify operational bottlenecks and generate actionable next steps. To extract unstructured knowledge, Herk showcases the grill-me interview skill, which prompts the user with probing questions about business priorities and saves the recorded responses directly into formatted Markdown documentation.
Finally, Herk applies Andrej Karpathy's LLM wiki concept, prompting Codex to crawl existing notes and build a linked, comprehensive wiki vault. He demonstrates the 3D brain skill, which renders the entire connected knowledge graph into a visual interface that can be run locally or hosted on a web domain. Addressing model configurations, Herk notes that while GPT-6 Astra excels at complex multi-step reasoning, smaller models such as 5.6 Sol or lower effort levels can perform routine knowledge retrieval while preserving weekly usage allowances.
The 4Cs Architecture for AI Operating Systems
Herk outlines a structural model for an AI Operating System consisting of Context, Connections, Capabilities, and Cadence. Context captures durable background information such as personal identity, corporate structure, and business objectives, while Connections link everyday tools like email, Slack, and project trackers. Together, Context and Connections form the second brain foundation. Capabilities define specific skills and tools, while Cadence manages autonomous agent schedules. Herk notes that building advanced capabilities and automations before establishing thorough context produces generic outputs rather than reliable results.
Local Project Setup and Repository Structure
Setting up the AI Operating System begins by creating a dedicated folder on the local machine and opening it as a project inside Codex. An onboarding skill guides the user through seven introductory questions, automatically generating directory trees, Markdown context files, and configuration files. Central to this structure is the AGENTS.md file, which mirrors CLAUDE.md for cross-tool compatibility. This file contains foundational system rules, voice preferences, and an explicit routing map that instructs the agent where specific categories of business data and reference material reside.
Maintaining Knowledge with Audits and Questioning Skills
To prevent knowledge from becoming stale, the system incorporates maintenance skills such as audit, level up, and grill me. Running an audit evaluates the repository against the 4Cs framework and produces a timestamped score and status report. The level up skill analyzes audit findings to suggest immediate automation priorities, while the grill me skill conducts in-depth interview sessions on specific business topics. Each interview automatically compiles answers into structured Markdown notes, systematically transferring tacit knowledge out of the operator's head and into the repository.
LLM Wikis, Visualization, and Model Selection
The system applies Andrej Karpathy's LLM wiki pattern to synthesize scattered files into an interconnected knowledge base. A 3D brain skill can generate an interactive visual graph showing the relationships among wiki pages, videos, meetings, and project files. When running these workflows in Codex, Herk demonstrates using GPT-6 Astra for heavy reasoning and multi-step agent creation, while noting that lighter models like 5.6 Sol or reduced reasoning effort settings can handle everyday retrieval tasks to conserve weekly usage limits.
The Bottom Line
The video establishes a practical architecture for building a persistent, file-based AI Operating System that turns frontier language models into fully informed operational partners. By grounding the setup in the 4Cs framework and using standardized AGENTS.md routing maps, it shows how to eliminate context switching and keep business knowledge organized locally. It leaves open the long-term technical trade-offs of hosting external tool integrations and managing compute quotas across growing enterprise teams.
FAQ
What is an AI Operating System and how does GPT-6 Astra power it?
An AI Operating System is a structured local file repository containing personal context, tool connections, executable skills, and automated cadences that an AI model references to perform tasks. In the video, GPT-6 Astra uses this system to access comprehensive business memory, write documentation, execute automated skills, and operate as an informed co-founder.
What are the four components of the 4Cs framework for an AI Operating System?
The 4Cs framework consists of Context, Connections, Capabilities, and Cadence. Context and Connections form the second brain by storing durable background knowledge and dynamic data feeds, while Capabilities and Cadence supply specific agent skills and autonomous scheduling.
What is the purpose of the AGENTS.md file in the Codex project workspace?
The AGENTS.md file defines the AI assistant core operating rules, formatting preferences, and routing maps. It instructs Codex where specific files, wikis, and reference documents live, and is read before processing user requests.
How does the audit skill evaluate the quality of an AI Operating System?
The audit skill inspects the repository across the 4Cs framework, checks the consistency of configuration files, generates a score out of 100, and writes a detailed markdown report outlining gaps and recommended next actions.
How does the grill me skill help build out context in an AI Operating System?
The grill me skill conducts targeted interview sessions where the AI asks the user questions about business goals, trade-offs, and priorities, automatically compiling the answers into organized Markdown notes in the repository.
Why does Nate Herk recommend using model 5.6 Sol alongside GPT-6 Astra?
While GPT-6 Astra is the most capable model for complex automation builds, using it for simple knowledge retrieval tasks can rapidly exhaust weekly usage limits, making 5.6 Sol a more cost-effective alternative for routine lookups.
Worth watching for
Developers, solopreneurs, and AI practitioners looking to build a structured, locally hosted second brain and autonomous operating system using Codex and frontier AI models.
- artificial-intelligence
- gpt-6-astra
- codex
- second-brain
- knowledge-management
- agentic-workflows