AI for Building a Second Brain Webinar June 11th
A short webinar that explains what an AI second brain is: your existing notes, emails, transcripts and CRM data linked by topic through an index. It also covers the three levels I use to build one.
Pour qui: Business owners and operators who keep hearing about AI second brains and want to understand how one works before they build their own.
À retenir
It is about your knowledge, not the AI
A second brain is 100% about your knowledge and 0% about which AI you use. The pieces already exist: Apple Notes, emails, transcripts, saved YouTube videos, Google Drive files and CRM records. They are already related to each other, but you don't know how. AI only finds and builds those connections, and you can plug any AI you want into the result.
The LLM Wiki method
Andrej Karpathy's LLM Wiki is the first version of this idea. An AI such as Claude Code reads every document, comes up with a list of topics, tags each document with them and links documents that share a topic. When a new document arrives, it reads it, tags it and adds the new links. That is the whole foundation of a second brain.
The index is the center
Once the connections exist, the AI writes them into one index file listing each topic and the links between documents. Underneath, the second brain is just text files with a very large index that connects to everything. Tell Claude Code to check the index of your second brain before it answers, so it finds the knowledge you already have.
Talk with your business
The value is being able to ask your business questions, like a digital double of your business or your clients. You can ask who a person is, what the status of the pipeline is, or how your content is performing, and then connect that back to your content research. You don't need a large model for this: I find Codex gives better usage for money, and it is enough for most business use cases.
Level 1: keyword links
The AI finds words repeated across documents and links them in the index. If two documents both say 'revenue', they get connected. The limitation is that a third document saying 'ARR' is skipped, even though it is also about revenue. Starting with level 1 is still very useful: it is enough to build and visualize the shape of your second brain.
Level 2: embeddings link by meaning
AI understands tokens and embeddings, not words. An embedding works like a dictionary that places tokens with similar meanings close together, like cat, dog and lion. With embeddings, the AI can connect a misspelled 'revenue', or 'ARR', to revenue. I created this level myself, and in my tests it adds new connections, but it is not fully there yet.
Level 3: split documents into sections
One recording can cover several subjects. A meeting with your boss might move from project A to theme B and then theme C. At level 3, another AI reads each document and splits it into sections by topic. Each section can then connect to other parts of your second brain, which creates even more connections.
Outils présentés
- Claude CodeThe AI that reads your documents, tags them by topic, links them and maintains the index (the approach used for the LLM Wiki).
- CodexOpens my LLM Wiki Visualizer during the demo. I recommend it as the best value for most business use cases.
- FigmaThe whiteboard I draw on to explain the three levels of a second brain.
- getdesign.mdA collection of working design systems. I recommend giving one to Claude as a base when building your own design system.
- Google GeminiIts Files API is suggested, together with Anthropic's, for building internal knowledge bases for agents.
- Genial Agency AI ResourcesWhere I share the second brain skill, the API pricing dashboard and my webinar recordings.
Questions du public
Do you still need RAG to feed your content to an AI agent?
RAG was used to chunk documents and feed them into the AI. Models now have very large context windows and a strong understanding of visual elements, so this is built into the models. To test it, take a long financial report that is available online (about 500 pages), give it to Claude and ask it precise questions. It will understand. In practice, you can put everything into the prompt.
What if I am building an agent that should only use internal information, not general search?
Look into the Files API from Gemini and from Anthropic, which are built for creating internal knowledge bases. I would add that the models are already good enough on their own. If you connect them to the right tools for your internal data, prompt them well and give them a good system prompt, they find the information well and fast.
Is the index useful for RAG and for setting up agents?
Honestly, I see very few people using RAG anymore. RAG architectures are rare because there are now simpler approaches that work better, and AI is good enough that you don't need complex setups. The index is useful because it makes sure your latest data is always available.
My design system is still off on color, size and spacing even after I feed in brand guidelines, code and screenshots. How do I fix it?
Check getdesign.md, which explains in detail what a design system is. I downloaded the design systems there and used them as inspiration for new ones. Give Claude a design system that already works, then tell it to use your new assets to change that design system or create a new one from it.
