The second brain

A second brain, as AR Labs builds it, is a set of plain-text files about a person's goals and work that an AI agent reads each session. With it, the agent knows your history, understands your goals, learns how you like to work, remembers things for you, helps you plan, and carries out some of the work. The files sit in one folder that an AI coding agent such as Claude Code reads every time it starts, so the agent works from your context rather than from nothing. VOX, the lab's working system, is a mature example, and we have taught the approach in a workshop for people who run their own businesses.

Chatbot or agent

A chatbot answers. It can give tailored advice and may remember a few facts, but the person still does the work. An agent acts: it reads your files, follows your methods, and uses tools to do things, such as putting a meeting in a calendar rather than explaining how to.

The difference that matters most is context. A language model's reality is whatever is in its context window when it responds. That context has four parts: system instructions loaded at startup, the current conversation, the tools it can reach, and persistent memory about you. A second brain is mostly the deliberate design of the first and last of these, which is context engineering applied to one person's life and work.

Built from plain files

The working material of a second brain is Markdown: plain text with light formatting for headings, lists and emphasis. It is readable and editable by a person, readable by an AI, and stored in no proprietary format. Claude Code and similar tools load an instruction file, called CLAUDE.md or AGENTS.md, from the folder they start in and from folders they work in, so each folder can carry its own explanation of what it contains and how to handle it.

A starter brain is small:

brain/
├── CLAUDE.md     the agent's instruction manual: who you are and how to work with you
├── GOALS.md      what you are working toward
├── working/      active tasks and notes
└── library/      reference material

The order of building matters. Everything starts with who the person is and what they want to achieve, and works down from there. Goals come first, because an agent that knows the goals can answer the question most task tools cannot: what is the most important thing to do next.

Capability levels

We teach the second brain as a progression. Each level builds on the one before, and how far a person goes tracks their own understanding of working with AI.

flowchart TD
  L0["0: Prompt and response, with memory in an instruction file"] --> L1["1: Custom slash commands for repeated processes"]
  L1 --> L2["2: Skills that package whole workflows"]
  L2 --> L3["3: A cognitive architecture of folders and instruction files"]
  L3 --> L4["4: Connections to external apps through MCP"]
  L4 --> L5["5: Automated workflows and sub-agents that work independently"]

• Level 0. Put yourself into the model's context with an instruction file: your name, your work, how you like to be addressed.
• Level 1. Ask the agent to turn a repeated process into a slash command, a saved prompt invoked by name. A first exercise is a command that interviews you one question at a time and writes the answers into your instruction file.
• Level 2. Skills formalise a sequence of actions into a reusable workflow the agent can load when needed.
• Level 3. A cognitive architecture: instruction files in each folder explain that folder's purpose, so the agent navigates the brain the way a person uses a table of contents. Loading the most general information first and detail only when needed is progressive disclosure, the same principle we use throughout our own system.
• Level 4. The Model Context Protocol (MCP) connects the agent to calendars, note-taking apps, email and other services, so it can act in them.
• Level 5. Long-running workflows and sub-agents that know what you want and work toward it on their own.

There is no shortcut through the last levels. Tools that speed up the mechanics keep appearing, but no tool can define for a person what they want, how they want things done, and what their preferences are. That work is the second brain.

Where it comes from

Our version draws on our own working system and on Daniel Miessler's Personal AI Infrastructure, a detailed public framework for personal AI assistants that influenced how we organised our workspace. The underlying discipline is context engineering, as described in Anthropic's and LangChain's guidance for building agents.

How we apply it

We created a workshop series, "Building Your Second Brain", for small business owners, freelancers, coaches and independent professionals. The first session, an introduction, covered the difference between chatbots and agents, context engineering, the capability levels and live demonstrations, and ended with homework: install Claude Code and a Markdown editor, create the starter brain, and build a first interview command that fills in a goals file. Materials for a deeper hands-on session, which works through guided prompts for a personal vision, roadmap and goals and installs a first skill, have been prepared.

Our own system, VOX, shows where the progression leads: instruction files throughout a large workspace, dozens of skills, connections to external services, and agents that carry work forward between sessions.