---
title: "What Is an AI Second Brain? The Four Cs Framework Explained"
url: "https://goremy.ai/blog/articles/ai-second-brain-four-cs-framework"
author: Luis Chavez-Mattos
editor: Luis Chavez-Mattos
editor_title: Director, Product
published: "2026-10-29T23:40:05.728Z"
updated: "2026-10-29T23:40:05.728Z"
topic: Agent Harnesses and Skills
description: "An AI second brain needs four parts: context, connections, capabilities and cadence. See how to build one with Claude Code and Obsidian, and audit each part."
---

# What Is an AI Second Brain? The Four Cs Framework Explained

> An AI second brain needs four parts: context, connections, capabilities and cadence. See how to build one with Claude Code and Obsidian, and audit each part.

Edited by Luis Chavez-Mattos, Director, Product · Published 2026-10-29

## What is an AI second brain?

An AI second brain is a coding agent that knows your work well enough to act like a teammate. It holds your goals, projects, people and history in files it can read, connects to the tools where you work, and runs repeatable tasks for you. The four Cs framework, credited to AI educator Nate Herk, breaks it into context, connections, capabilities and cadence. If all four work, the underlying model matters much less.

## TL;DR

- A second brain should **know your business off the top of its head**, without digging through memory for basic questions.
- **Context is the foundation**: a few markdown files loaded every session, backed by a larger Karpathy-style wiki.
- **Connections** let the agent work where you work (email, tasks, calendar, code, chat) instead of being a smart chatbot in a box.
- **Capabilities** are the repeatable jobs you hand off, usually packaged as skills.
- **Cadence** covers how and when the system works for you, including reaching it remotely, and the framework doubles as a **checklist for auditing** an existing setup.
- If an audit fails, tell the second brain what felt wrong and **fix it together** with the agent.

## Why use a framework instead of piling on tools?

The creator who presented this framework argues that people overcomplicate their second brains. His own stack is small: Claude Code as the agent, Obsidian for the knowledge vault, an Excalidraw plugin inside Obsidian for diagrams, OpenRouter for images and other models, and a handful of integrations. He says it fits on one slide and runs on "a few tools and a folder of markdown files."

He credits capable coding agents for that simplicity. The agent runs commands directly on his computer, so he does not pay separate services for tasks like diagrams or presentations. He reports that the system saves him roughly 25 to 30 hours a week, up from about 10 when he first built it. Those are his own figures, not independent measurements.

The framework is model-agnostic. He says he has run the same system on different Claude models, on Codex, and on a Kimi model for testing, and it kept working. His conclusion: the context determines how well the system works, not the model.

## What is context, and how do you build it?

Context is what the agent knows before you say anything. The setup described uses two layers.

**Layer one: core files loaded every session.**

| File | Purpose |
|---|---|
| soul.md | How the agent should behave and act |
| user.md | Who you are and the core facts of your business |
| memory.md | The most important memories, promoted from the larger knowledge base (such as main business objectives) |

A Claude Code hook loads these at the start of each conversation, alongside your global rules file (CLAUDE.md). The files also describe how the deeper knowledge base is organized, so the agent can search it efficiently.

**Layer two: a Karpathy-style LLM wiki.** This is a knowledge base of plain markdown in Obsidian. Every conversation produces a summary that is ingested into daily logs. The wiki also keeps an entity document for each thing in the business: each video in progress, each company worked with, each team member or open source collaborator. Those documents are maintained over time.

Markdown keeps the whole thing portable. You can sync the vault across devices, read it yourself, and let the agent search it with ordinary file tools. Obsidian is the choice here, though the source notes that something like Notion could work.

### How do you audit context?

Open a fresh session and ask what your business does and who works there. A good second brain answers like a teammate, quickly, without a deep search through memory or your platforms. If it stumbles, say so to the agent. Something like "it doesn't feel like you know my business off the top of your head, what can we do about that?" starts a useful conversation about what belongs in the core files.

## What are connections?

Connections are the links between the agent and the places you work. The source argues that context matters most, but context alone gives you "an ultra smart chatbot." It cannot triage your inbox, check your task list or file a document.

The example setup connects to:

- Google Calendar and Gmail
- Asana for task management
- GitHub repositories
- Google Drive for files
- Slack, which he uses to talk to the agent from his phone and to help manage team communication

Most platforms offer an API or an MCP server, so adding one is usually light work. Where no easy MCP server exists, the creator builds a small command-line tool of his own and has the agent use that.

### How do you audit connections?

Ask for a Monday morning catchup: your priorities, what needs a reply, who needs your attention. If the agent cannot pull that together quickly from across your platforms, a connection is missing or weak. Each gap is a to-do item for the next integration.

## What are capabilities?

Capabilities are the repeatable jobs your second brain can do. In a Claude Code setup these are typically skills: saved instructions, and sometimes scripts, for work you do again and again. The creator mentions skills as part of how he leans on his agent to make improvements to itself every day, though not fully by itself, and notes that the agent runs local commands to produce things like diagrams, presentations and video edits.

The source material available here does not walk through the capabilities audit in detail, so the following is a practical reading rather than the creator's checklist:

- List the tasks you repeat weekly (research, content prep, reports, diagrams).
- Check whether each one is a skill the agent can run on request.
- When you catch yourself explaining the same process twice, ask the agent to turn it into a skill.
- Review skills periodically and ask which ones are working and which are not.

That last habit matches what the creator describes: regularly telling the agent what is working, what is not, and asking how to improve the context system and skills.

## What is cadence?

Cadence is the rhythm of the system: when it runs, how often it updates, and how you reach it. The source text cuts off before covering this C, so treat the following as general interpretation tied to details that were mentioned.

Some cadence is already visible in the setup described. Every conversation is summarized into daily logs, so the knowledge base refreshes continuously. The creator also runs his system on a cloud VPS, which he says costs about $20 a month (he names DigitalOcean, with Hostinger as a cheaper option). That lets him message it from his phone through Slack even when his computer is off. He stresses that hosting is optional and that a local build is the right place to start.

A reasonable cadence audit asks whether the system does anything without being prompted, whether its knowledge stays current without manual upkeep, and whether you can reach it where and when you need it.

## How do you audit your own second brain?

Run the tests in order, since later Cs depend on earlier ones.

1. **Context:** In a fresh session, ask what your business does and who works there. Pass means a fast, teammate-quality answer.
2. **Connections:** Ask for a Monday morning catchup. Pass means it finds priorities, pending replies and people who need you across your tools.
3. **Capabilities:** Pick a task you repeat. Pass means a skill handles it without re-explaining.
4. **Cadence:** Check that memory updates itself and that you can reach the agent when you are away from your desk.

When a test fails, do not rebuild from scratch. Tell the agent which test failed and work out the fix with it. The creator says this is how he improved his own system, though he notes the agent does not improve itself fully on its own and he stays in the loop.

## Does the harness or model matter?

Less than you might expect. The source says it is easy to switch between harnesses and models, though you should not do it often. Files in markdown, integrations through standard APIs and MCP, and skills stored as instructions all travel with you. The creator also says the four Cs apply whether you build on Claude Code or on other agent frameworks such as OpenClaw or Hermes.

## Frequently Asked Questions

### What do I need to start an AI second brain?

A coding agent (Claude Code in the example), a folder of markdown files, and a few core documents such as soul, user and memory files. Obsidian is a convenient way to view and edit the vault. You can build locally first and skip cloud hosting.

### What is a Karpathy-style LLM wiki?

It is a markdown knowledge base maintained by an LLM, with entity pages and logs the agent can search and update. In the setup described, each conversation is summarized into daily logs, and entity documents track videos, companies and people.

### Do I need a specific model for this to work?

No. The creator says he has run the same system on several Claude models, on Codex and on a Kimi model, and that the context you provide matters more than the model.

### What should I do if my second brain fails an audit?

Tell it directly. Describe what felt wrong, such as not knowing your business or missing a platform, and ask what to change. Then work through the fix together, whether that means editing core files, adding an integration or writing a skill.