Finally built a personal knowledge system that will actually increase my productivity: second brain + AI generated learning vaults with local LLM
I am relatively new to using Obsidian. I worked on building a system for myself over the past 1 to 2 months and I finally have something I believe will genuinely help me and not just be a gimmick. My structure:
I have a main vault that I call my second brain. This vault only has information that I type in myself and only about information I care and would like to come back to in the future to easily recall/study/refer to. I use claudecode with this but only to query my vault (no editing whatsoever!). I use a claude md file and multiple skills to help me make the best use of the information in my vault.
I also have other "learning vaults" that are modified versions of Karpathy's LLM wiki. I've created a successful Claude md file that is able to give AI clear instructions to read my class notes and create a vault. This vault follows the same structure as my main vault to make it easy for me to query and I personally also think it is more efficient than other systems I have seen. With the help of AI, I am creating the vault which I can access through obsidian to study and also query the AI directly to learn from the notes as well.
I am a college student right now and I want to do a PhD later and I am currently working in a lab so I read a lot of research papers. I am working on creating a claude md file for learning vaults focused on extracting useful information from research papers as well.
I am big on privacy so I use a local LLM with claude code. The model I am currently using is qwen3.6:35b-a3b-coding-nvfp4 with Ollama. I have a MacBook M4 Pro with 48 GB RAM.
Also, a cloud based AI isn't really an option for my learning vaults anyways because ingesting my notes and building a vault uses way too many tokens and easily crosses the claude pro subscriptions's limit.
This basically sums up the structure of my vault:
"The vault is called Brain and uses a strict three-level hierarchy: Topics → Subtopics → Notes
The hierarchy is strictly one-to-one:
- Each note belongs to exactly one subtopic via its
Categoryproperty - Each subtopic belongs to exactly one topic via its
topicproperty - Do not look for multiple parents — every file has exactly one
- Notes link to other relevant notes through internal links
Tags
The tags property can contain two types of tags. Do not confuse them.
Organizational Tags (determine note type and folder)
Each note has exactly one of these: article, author, book, documentary, explainer, item, log, movie, people, podcast, process, research paper, technique, video
Special structural tags (not content notes): subtopic, topic
Personal Query Tags (filtering only, not structural)
A note may optionally also have one of these: young, adolescent, adult
This structure is important to me to help keep things easy to find and also make the best use of the local graph view mode. I feel a local graph view mode of a topic and subtopic is very very helpful.
I am sharing this just out of excitement from having completed these finally. I have tried to keep everything very brief. Happy to explain things in more detail and open to suggestions!
