Summary

I set up Obsidian as a place to organize all of my projects plus my learning and career pivot. This became the foundation of the AI-supported Second Brain and Personal Assistant.

🤖AI Transparency Statement: This article was fully written by me, then went through basic grammar checks and a non-editing feedback loop on clarity with Claude Sonnet 4.6.

About the Second Brain Series

This is the first in a series showing how I’m implementing AI and LLMs to manage my workflows and knowledge capture or recall. These articles aren’t intended to be a step-by-step guide. You can find endless sources of those elsewhere. The idea is to demonstrate one of the ways I’m using AI and LLMs to accelerate my goals and projects.

Background

I’ve evolved my note-taking over time. What was originally hand-written sticky notes or emails to myself to not forget something changed into primarily dedicated hand-written notebooks, supported by some digital note-taking in OneNote and still a lot of context spread between emails, Miro virtual whiteboards, random docs, and more. It was nearly impossible for me to quickly recall anything from the past, and increasingly frustrating when I knew I had logged something, but couldn’t find it again. And the recall could be even worse when working with peers, even when using other systems like Confluence or Slack Canvas.

After leaving my last role I was faced with a new challenge to manage a career pivot, learn lots about new areas like Climate and AI Safety, how to use AI and LLM, and how I wanted to integrate those into personal and public projects, I decided to look for a better solution. After some searching, I landed on Obsidian.md for a few reasons: free, simple markdown files (also preferred by LLMs), non-database, private, and community powered.

Getting Started

The fun thing once you start reading about how to use Obsidian is that there is no right way. Do you build out a folder hierarchy? Or leave everything flat and use tags…or use extensive linking? Organize by project or source? And this gets even more interesting when you later arrive at trying to integrate AI into this picture, which I got hung up on again.

The best thing though was just to get started, and figure out what worked for me. Most people would describe me as highly organized, and maybe that’s what led me to a more formal folder hierarchy. Still though, this has continued to evolved over time. For example, one thorny area was trying to separate my general learning on AI and LLMs, which includes frontier companies, models and tools from my interest and projects to get Local LLMs up and running locally. Initially separated, they ended up merging back after establishing a lot of confusing overlap. But, even now, the local LLM aspect sits at odds with the larger ‘homelab’ project, which the implementation will fit into.

Long story short? I found it better to just start, accept imperfection, and be willing to re-factor once the pain-level becomes big enough. Some things still uncomfortably live in one place when maybe they could merge into another.

What I’m Capturing

As of this writing I have 4 high-level folders (AI and LLMs, Jobs and Career, Homelab, and PaulsForge). Each of those has 5-10 layers below them, capturing a core sub-theme. And then the next layer covers things like companies and organizations, interesting articles or webinars, or project ideas or more. That final layer structure really starts to vary depending on the topic-area.

To help keep things connected, I do try to create cross-links and have started some limited use of tags, although I’m not as consistent with tags.

Key Challenges, Workarounds and Ideas

Beyond folders, the Obsidian vault has continued to evolve as the amount of content grows, and my interests and ideas shift and grow themselves. This has presented some challenges.

Clipping

I’m constantly finding interesting companies, programs, technologies, how-tos and more. Sometimes, I don’t have time to review at all and I just want a quick capture. Other times, all that makes sense is capturing a little bit of context or my key takeaways for future recall. And then sometimes I want to capture a full article or podcast (transcript) because I want to capture the full context for reference later, and potentially add deeper notes into it.

Obsidian has a popular Web Clipper plugin that basically dumps a web page to markdown. This is great, but I can’t clip every interesting thing I read. This is an organizational nightmare, and without a good recall or search system is flooding Obsidian’s basic text search or even a simple LLM scan with way too much context (tokens).

Workaround: For now I’ve had to balance clipping the articles that have struck me the most with trying to build a consistent system to dump links with short summaries into their right folders.

Ideas: Ideally I could intake everything interesting easily, automatically classify the level of detail I want to capture, and then on top of that layer some kind of agentic summary processing to keep a middle ‘wiki’ layer that’s easier for myself and search, but still inks back to the full content as needed.

To-Dos

Still largely unsolved. Often times I want to flag a task or something I want to read or review later. I don’t have a great system for these yet except to separate a file into 2 sections of Read or Completed vs Unread or Not Completed.

Workaround: I’m often still using the tried-and-true method of emailing myself a link I want to make sure I read soon.

Ideas: Similarly, an intake system with proper classification an management, likely agentic. Can re-surface proactively at good times, whether something is due, or its just a good time to consume something.

Searching

The default Obsidian search is just a basic full-text search that you can layer some limited logic on top of. As the amount of content has grown, and my memory of the exact keyword I might want to hit has grown fuzzier, it’s become challenging at times to track things down. I’m more successful drilling down into folders than using the built-in search.

Workaround: Right now my strong folder hierarchy is mostly saving me.

Ideas: Agents, vector search, RAGs, LLMs…and more! Paired with some summarization and other tricks to improve recall and reduce tokens.

Where This is Going

For this vault to be effective and efficient to use, it needs more intelligence. And this doesn’t just reflect my personal needs, but also the pains I felt in corporate of having to constantly help dig up and recall past work and decisions. I think there’s a lot of interesting applications for this work in my next roles where I can implement a similar system within my organization’s toolsets. The core ideas is to log feedback, requests, decisions and their logic, requirements, research, other interesting articles and more. I can do this in a way that makes sense to me, then 1) automatically have the content processed down and maintained into summaries or wikis, while retaining the raw content as reference; 2) make it easy for more intelligent search tools to query against either.

In the next parts of the Second Brain series, I’ll start sharing more about what I’ve tried and learned as part of that process and continue to plot out a path to a truly AI-forward way of approaching my work.