Microsoft Cloud IT Pro Podcast

The episode contains substantive technical content about building LLM wikis with Obsidian, including concrete workflow examples (Web Clipper integration, RSS dashboard plugin, Python automation, Copilot Cowork scheduling). However, significant portions are occupied by tangential discussions about local AI models and resource contention that dilute the focus, and some explanations lack depth (e.g., brief mention of token optimization without detailed metrics or trade-offs).

“So you take this gist, and you put it into your LLM interface, in this case, let me say, like, Claude code. Just paste it straight into the terminal and add a line at the end that says, I want you to turn this into a context file for this vault”

“So I actually created a template dot m d file, and then I went back and I modified my cloud context. And I said, okay. Hey. When you ingest things, make sure that the outputs use this template.”

The episode applies existing LLM wiki concepts (sourced from Andrew Karpathy's gist) to practical IT pro workflows, which is somewhat original contextualization. However, the core idea and architecture are not novel - this is a direct adaptation of published frameworks. The originality lies in execution details (combining RSS dashboard, Copilot Cowork scheduling, email-to-markdown extraction) rather than fundamental invention. The speaker acknowledges this is built on prior work.

“So Andrew Kapathy, who's kinda, like, one of the original or or original researchers over at OpenAI, I believe he's also the person who coined the term vibe coding. So he's kinda pretty close to it. He's been in this space for a little bit?”

“I'll have a link in the show notes for everybody to a gist that he posted. And and this gist is just an LLM Wiki.”

Scott is an experienced IT pro and practitioner who has actually built and iterated on LLM wiki implementations, demonstrating hands-on expertise. However, the episode features only one guest (Scott being interviewed by Ben), and while Scott shows competence, he is not a recognized authority figure or someone operating at massive scale. He is a skilled practitioner sharing real workflows, but lacks the seniority or outside credibility of a C-level operator or recognized thought leader at a major organization.

“I went and iterated over mine a little bit. So I I because I was using Obsidian Web Clipper uses a template when it loads data in”

“So over time, I've been on this journey of saying, hey. All the stuff that you're doing that I see you're repeating yourself doing, but it's you as, like, the LLM thinking about it, just go slam that in my script.”

The episode includes concrete technical specifics: named tools (Obsidian, RSS Dashboard plugin, Copilot Cowork, GitHub Copilot CLI, Cloud Code, OpenRouter, AntiGravity, Claude Max 5x), actual file structures (clippings directory, Wiki/sources, Wiki/concepts, Wiki/people), specific commands (brain.py ingest, brain.py lint, brain.py organize), and real token-burning examples (five-hour quota burned in 12 minutes, AntiGravity tokens exhausted in six days after two ingests). However, lacks named organizations, specific metrics on vault sizes, exact cost figures, or measurable performance improvements.

“I created me a Python script called brain dot p y. And off brain dot p y, it created a number of subarguments. So it took my ingest flow, and after, like, three ingests and it had all the context, and it kinda knew what it was doing repeatedly, it built me a brain dot p y, slash ingest or, like, dash ingest command.”

“I did mine. I did the podcast. It was granted, it was 400 and, what, 28 episodes. I had all the transcripts, so it was some big text documents, like thirty minutes of talking turns into a lot of text. And I was using Claude, and I was doing the Claude Max five x. So this is the $100 a month plan on Claude, and you get, like, a five hour quota. When I started ingesting podcast episodes, and even I experienced this even. I tried to put a bunch of stuff in Python, do as much as I could still burn through my five hours of quota in, like, twelve minutes.”

Ben asks some clarifying follow-ups (e.g., "Have you tried it the other way", "Have you compared the LLM that gets built by Cowork versus Claude", questions about throttling and quotas) and shows genuine curiosity about cost and security implications. However, the conversation lacks depth - follow-ups are often brief and don't push back on claims or explore contradictions. When security concerns (IRM, MIP, sensitivity labels) are raised, the hosts acknowledge limitations but don't probe deeply. The host also frequently allows Scott to steer topics (tangential discussions about local models take significant time) without redirect.

“Have you tried it the other way where once you have your Obsidian Vault and you're syncing it down with OneDrive, you save articles into your raw folder in OneDrive and have Copilot Cowork actually ingest all the stuff you're saving locally”

“Have you compared the LLM that gets built or the Wiki that gets built by Cowork versus a Wiki that gets built by Claude just to see side by side how”