Episode 429: Getting started with LLM Wikis

- Welcome to episode 429 - of the Microsoft Cloud IT Pro podcast recorded - live on 05/29/2026. - This is a show about Microsoft three sixty - five and Azure from the perspective of IT - pros and end users, where we discuss a - topic or recent news and how it relates - to you. - In this episode, Ben and Scott dig into - the concept of LLM wikis, specifically - building - personal knowledge management vault using Obsidian, - markdown, and AI tooling like Cloud Code, GitHub - Copilot CLI, and Copilot Cowork.

- Scott walks through how he wired up Obsidian - Web Clipper and an RSS dashboard plugin to - feed articles into his vault automatically, - then had the LLM help build a Python - script to automate the ingestion workflow and cut - down on token usage. - Then he expands into how Copilot Cowork fits - into this workflow as a scheduling harness with - practical examples of using it to pull email - from an inbox daily, convert messages to markdown, - and generate a prioritized to do list.

Let's - dive into the world of LLM wikis. - I have a theory. Windows? No.

I mean, - yes. I run too much stuff. Yes. But - no.

Well, I think we all run too - much stuff. I've noticed lately that and I - think I talked about this a couple weeks - ago, the whole I've got cotypist on my - Mac doing type ahead prediction - with with a locally downloaded model. - I've got - Type Whisper - running in the background with a model loaded - in case I just wanna do text dictation. - Like, I I I I was thinking about - it the other day.

At any given time, - local model running, I probably have at least - four instances of the same model or a - similar model all loaded up because these apps - don't have, like, a common cache interface. Like, - they don't use, like, hugging face cache or - things like that. They all download, like, locally - into their packages and their directories. - So at least - for me, I think a lot of the - things that I'm running into lately are actually - resource contention.

Like, I haven't materially changed my - workloads - or my workflows, but I have a bunch - of the bunch more of these apps running - in the background that I didn't have before. - I was trying to do a little bit - of an audit, and it's not driving, like, - crazy memory pressure, crazy CPU pressure. - But, like, when some of these models load - and unload, like, you've got, like, I don't - know, like, Gemma four sitting there ready to - go. It's like, oh, hey.

I had four - gigs of RAM that just got evicted for - this thing over here, and then it's gotta - reload rapidly for this like, it I don't - know. That that would be my guess is - we're all running, like, so much, like, local - AI slops that we don't even know it. - So little of it we can control. Like, - I don't remember if we talked about this - or when this came out, but Google Chrome - now downloading, what, a four gig - LLM model in the background as part of - the browser?

The Gemma model in the background. - Yep. Yep. So that same model is the - same model that I'm using in Cotypeist, which - is - similar to what I'm so so, like, I've - got this app called Type Whisper.

- Anyhoo. So Type Whisper is one of the - dictation apps. But rather than using, like, Whisperflow - or Super Whisper, - everything - processed locally on device, you can choose, like, - hey. I I wanna use Parakeet.

But the - cool thing it does is it does - live processing, like live transcription, - and then it does post processing as well, - which is nice. So you can do that - thing where you're just saying, hey. I'd like - to make a list. Number one is this.

- Number two is this. Number three is this. - And, normally, that would just be a big - flat wall of text going across. But if - you post process it and it gets smarter, - it actually renders, like, one, two, three, and - four into a list, and that's the text - that is expressed that you just dictated.

- But I was thinking about that one the - other day. I was like, oh, this is - getting slower and slower. And I went and - looked, and I turned on, like, the pre - integration, - the Run It thing, the post integration. - So that one had to get kicked to - the curb.

It was nice, but - it was it was slowing things down for - sure. Yeah. - All these things. - You know what other things well, nope.

I - won't go down that rabbit hole. We can - talk about other things. But speaking of more - things we're running, you and I both started - playing with something within the last I don't - know when you started. I started, like, a - week and a half ago - playing with more things to run on our - computers.

- Yes. And this is not maybe we should - preface this. This is not specifically, I would - say, Microsoft or March - Microsoft three sixty five or Azure episode. However, - it incorporates into our work with Microsoft three - sixty five and Azure and how we're starting - to track some of our resources.

I have - a tie in I have a tie in - for you there. I'm actually gonna yeah. Do - you have a tie in for me? Yep.

- Okay. So if you're listening and you're like, - there's not gonna be a tie in, Scott - has a tie in. I don't know when - it's gonna drop. So you better listen to - the whole episode - to figure out the tie in.

Stick around. - Okay. I've been playing around - and been watching all these just videos fly - by and articles fly by - around the concept - of LLM wikis - and - leveraging - LLMs - to - manage a wiki resource for you. And, specifically, - like, I've been doing this with Obsidian Vaults.

- So Obsidian Vaults work very nicely here. They're - marked down plain text. LLMs process plain text - pretty darn well. They understand markdown pretty well.

- It's kinda, like, a a de facto standard - for authoring these things. Right? All the skills, - agents, you write everything. Everything you do is - in markdown.

- So an LLM Wiki, if - you have a compendium of knowledge that you - wanna put together, and that could be around - anything. It could be maybe around a topic - that you're researching. - So I don't know. Let's say you want - to - in M 365 Land, do you want to - go explore the current single the current licensing - constructs and compare this e SKU to this - business premium SKU or to this add on?

- You could go and take, like, all those - pricing pages and all the documentation around that, - and you could put it in just a - big dumping ground of markdown, just a raw - source. And then you can have an LLM - reason across that data - and extract information around it. So understand - what was the source data that came with - it, maybe what are the general concepts around - it, entities, - all these different - sorts of things that are out there. - 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? Been in this space for a - little bit. Was recently hired by Anthropic - post - getting out there and putting some of these - things out.

But 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. - So you go and kinda read through this - thing, and it's just llm-wiki.md.

- And - the core idea - of this whole thing, like I said, is - you kinda have a big dumping ground. - You could think about it as maybe, like, - a folder that you called raw or you - called ingest, - you called it clippings, - something like that. - And what you do is you keep putting - articles in there, reference materials, - but you put it in as markdown. - So this whole thing just becomes - a big bank of markdown that's sitting out - there.

And - you direct the LLM that anytime - a new thing shows up in my raw - directory - to reason over it and then extract that - information, extract entities, extract core concepts, tell me - what the source of it is, and put - it all together. - And this is the nice thing about something - like Obsidian. So Obsidian, a personal knowledge management - tool, PKM, - But Obsidian - does - Wiki style links, so you can have all - these things start to relate to each other. - And then you get the kinda graph based - views that that come along with that as - well.

- So it takes almost nothing to wire it - up, first of all, to get started. So - it's pretty much go out and take this - gist, - and, yes, you need to have access to - some kind of LLM interface. Now that LLM - interface - could be hang with me here. Could be - quad code, could be Gemini CLI.

Well, actually, - Gemini do CLI just went away. Anti anti - gravity, but there's also anti gravity CLI. - It could be GitHub Copilot CLI, codex, - whatever that is. Or if you have access - to, I don't know, say, Copilot Cowork - and in in and the Frontier agent because - you're in the Frontier program, - you can also run an LLM Wiki inside - of Copilot Coburg.

Have you tried this? I'm - curious to see. Did and it works? Oh, - yeah.

It opens up a whole interesting world - of possibilities here. Hold on. Stick with me. - Now you have me curious.

- If you're managing file storage today, you've probably - felt the pressure. Refresh cycles are harder to - plan, costs and complexity are increasing, especially as - file workloads scale across regions. - That's why Microsoft customers trust Nasuni. As an - Azure certified software, - Nasuni provides a global file system designed for - enterprise scale environments, keeping file data secure, governed, - and accessible across Azure environments without added infrastructure - complexity.

- It's the foundation to keep your unstructured data - stable even as your infrastructure evolves. To see - how much you could save while keeping your - initiatives on track, visit nasuni.com/tco. - That's nasuni.

com/tco. - So let's talk about wiring all this up. - So so you basically go out the and - I've done a couple of these now, like, - and been kinda playing around. And you've been - playing around with one to build, like, a - basically, an Obsidian Vault for us around the - podcast.

We can go back and say podcast - stuff in. It's fascinating. When did we talk - about this? On what date this happened?

How - many times did we have this guest on? - And now you've got all that information in - text in one place. Show me all the - episodes where we talked about a specific topic. - Like, it's been really interesting.

Go show me - what topics. We've pulled analytics into it. Go - show me what topics - tied into - analytics. Yeah.

All the things. Go take this - gist, and you can literally copy and paste - it. I don't know if this is how - you started with the one that you worked - on. I'm curious how you started, and I'll - tell you how I started.

So I took - this gist, and I copied and pasted it - straight into a CLI. I took it and - put it right into actually, you know what - I did? Because I wanted to do this, - like, a freeway that anybody could get started - with. So I did it in cloud code, - but I did it in cloud code tied - to - OpenRouter - and the free models in OpenRouter.

So I - wasn't using a paid anthropic or cloud plan. - I wasn't using, like, pay per token. I - did all this stuff for free. You could - do it free for free with anti gravity - and things like that.

It gets a little - harder to do free later, like, once the - vaults get a little bit bigger, but you - can totally start out Really hard to do - free later. But you can start out for - free. 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, or, like, - I want you to create the second brain - for Scott that manages - everything he finds on the Internet wants to - read about later.

Right? So you give it - a little bit of context. That's it. You - just put in that one little sentence at - the end, fire it off.

It will create - a Claude in the case of Claude, it - will create a claude dot MD. You did - this in GitHub Copilot. It'll create you, like, - a Copilot instructions dot MD. I recommend doing - a claud dot MD as a context file, - and then you can just pull it from, - like, LLM to LLM because all the various - or, I guess, from agent to harness to - harness.

Because you can take a claud dot - MD and use it in claud. You can - take a clog dot m d and use - it in anti gravity CLI. You can take - it using GitHub Copilot CI. So so that - context file becomes very portable.

So you do - that, and that's where you - start. And then it's gonna create your, in - my case, a clog dot m d. - And in my Cloud dot md, I said, - okay. I'm gonna go do a little bit - of light iteration here on some of the - instructions - that that I want this thing to have, - right, and have for its context.

- So that was I've got a root folder - called clippings. - I believe in the original gist, it's called, - like, raw. I don't know. It didn't sit - well with me.

I was like, yeah. I'm - gonna put Obsidian clippings in here, like Obsidian - web clipper, so I call mine clippings. And - then I added a couple different Wiki directories, - so it automatically does one for it does, - like, Wiki slash sources, - entities, Wiki slash concepts. - I wanted Wiki slash people because I knew - I was gonna be putting things in there - about, like, the podcast and you and I, - so I wanted it to extract, like, Ben - and Scott or Ben and Scott together, - kinds of things.

And then - all you do is go out and start - putting some markdown in that in my case, - my clippings directory. So the way I did - that is with Obsidian Web Clipper. So Obsidian - Web Clipper is available on every platform. It's - there on iOS and Android, so you can - use it on your mobile.

It's there - on Edge and Chrome and Safari. - So any kind of thing that you end - up looking at on the web, all of - a sudden becomes a resource that you can - take and you could say, hey. I wanna - reason over this later in my Wiki. So - maybe you start small.

Go grab, like, five - articles that you're interested in and put them - in there. And then all you do is, - say, ingest my clippings or ingest my raw - stuff, depending on how you set up that - directory and how you framed out your ingest - flow. And then it'll go, and it'll just - build this stuff out. 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 I wanted - all my reasoned over data to also follow - a similar template. So it had a title, - a publish date, a create a date, tags, - reference resources, things like that. 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. Super easy. Right?

- So and you just start right there. - And - you start to get this little bit of, - like, this is interesting. - And then you start wondering what else you - can do with it. But it runs away - from you very fast because - you're ingesting text and maybe not just a - couple lines of text, but let's take that - m $3.

65 pricing example. Right? I I wanna - build something that can go reason over that. - You're going and grabbing every pricing page, potentially - the images off those pages and doing, like, - OCR and computer vision over them.

So it's - pages and hundreds of kilobytes of text that - you're putting in, which is heavy on the - token side, both on input and output. Yes. - So this is where something like maybe like - OpenRouter and a free solution - starts to run away from you a little - bit. So I did OpenRouter.

I got by - with maybe, like, my first 10 clippings in - doing this. And when I ran out of - stuff in OpenRouter, I was like, well, I - just wanna see, like, how far I can - push this free thing. So I went over - to Anti Gravity, which was recently released, and - that uses Gemini and Gemini Flash under the - hood. Let me tell you, anti gravity burns - some tokens.

I think I burned through all - my tokens in anti gravity, like, on my - after, like, my first two ingests, - and it burned them for, like, six days. - You have no more usage for the next - six days. - But what you do there is - LLMs, - they really like to overthink things. Right?

And - so they'll start to do things like, you'll - notice these patterns when, in my case, I - was saying ingest my clippings. - I noticed every single time, no matter - which harness I was using or which model - I was using, all of them were doing - similar things. They would say, like, oh, well, - I need to go figure out how to - list your ingest directory. So every single one - was saying, thinking thinking thinking.

Okay. I'm gonna - list your your ingest directory. So I'm gonna - go do, like, an lsstar.md - kinda thing.

You can automate all that stuff. - You don't have to have the LLM do - it. So the trick is you come back - or you or you don't have to harness - and the model do it. So you come - back and you say, hey.

I've got this - context file now, and we've come up with - a workflow. How can I automate this workflow? - Can we make parts of this workflow scriptable? - Just ask the question.

- And in my case, like, the one that - I've been really living in, I called it - my brain bank. I said, hey. This is - Scott's second brain. - So it 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. And then - I said, oh, hey. You're doing this.

Why - don't you build me, like, a like a - linting command to make sure that you're always - matching up to my template? Because I used - to use spinning a lot of tokens trying - to go back and read that template dot - MD and then get it back rear. So - it made me brain dot p y lint, - made me brain dot p y organize, - made me brain dot p y, - clean up any old clippings, maybe, like, organize - them into subfolders, - things like that, link them back to sources. - 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. Just keep putting it in there, putting - it in there, putting it in there. So - then you're not burning all the tokens. Like, - it doesn't have to and then you tell - it, go update my context file.

So if - I look at how I started from, let's - say, like, my ingest flow in my claw.md, - originally, - that was - something like - go - it said, hey. Your vault structure for clippings, - things like that. So my operations for ingest - was, like, step one, go list this directory.

- Now step one has turned into research. - Run Python three scripts - brain dot p y scan ingest, - and then - take that output and then go check this. - Go run brain dot p y again, get - this list output, compare those two things. So - that way, it doesn't have to, you know, - start immediately every single time enumerating a whole - directory and running over it - and being very heavy.

So you can start - to very quickly, like, drive it down into, - like, scriptability and less and that was the - part that I kind of enjoyed. Right? Like, - IT pro hat. Like, hey.

Let's go get - this going. I think you were doing some - PowerShell scripts and things on your side, - on the Windows side. But you just kinda - keep going down that path, going down that - path, going down that path. And then - because you're doing this all in an Obsidian - markdown style, now you go and open it - up in Obsidian.

Just go open your project - directory and what you're doing in there, - and you'll get access to - all the features of Obsidian. So that includes - tagging. It includes searching. It includes the PKM - with the knowledge graph.

- Obsidian's - like a canonical markdown editor. It's really darn - good at it. You and I had been - chatting about one of the ways we do - research for the podcast is by following a - bunch of RSS feeds. What's the easiest way - to ingest the RSS feed?

Well, I've been - going through, like, Feedly or Reader and opening - it in a website and then hitting the - Obsidian web clipper and doing things like that. - There's this really cool plugin for Obsidian called - RSS dashboard where you can just follow an - RSS feed. And as you follow, it's - it and now because it's an Obsidian vault - and it's just marked down, and I can - have it in my OneDrive - or you can have it in your Google - Drive or whatever, - It's actually kinda like Feedly, and that suddenly - it's a cloud sync service.

Right? It's following - you around in your configuration. - But RSS dashboard is really cool because every - feed article that comes through, you can just - click a red or an unread button to - to flag it or get rid of it. - But if you click save on it, it - has a really nifty templating system.

So I - was able to take my template dot m - d, which, again, was just based on an - Obsidian web clipper kinda default template. But I - was able to make that my template for - RSS dashboard - and have it every time I saved an - article from RSS, - it saves it as markdown over in my - clippings directory automatically. Right? Like like like, yes, - I have to click save on the article, - but it gives me this kind of, like, - one stop shop - ingest flow to get those things kind kinda - through.

- So really interesting way to go and kind - of reason over personal data, - really any kind of project that you're doing. - I I would recommend constraining this stuff. Think - about it maybe like, if you're coming from, - like, - if you're coming from Claude Land, maybe think - about it as, like, project based. Or if - you're coming from, like, the cowork side of - the house, - maybe you use, like, notebooks in cowork, things - like that.

Like, let me group these documents - together so that I can go do some - research on them kinda thing. I do find - that, like, I I I think that they - do a lot better when it's kind of - same knowledge area to reason over rather than - having a whole bunch of different things there. - Because, again, you could see eventually, you get - back to, like, really heavy context windows and - and things like that that happen along the - way. But so let's talk about now hey.

- Great. You did this in our RSS feed. - Or, alright. You told me that, like, I - could dig an article from the web, and - I can clip it in here.

Sure. Whatever. - Copilot cowork. - And Copilot cowork, - the Frontier agent.

So the Frontier agent that's - out there has custom skills. So you can - take - the very same - kind of flow - around - having a context, - in this case, like, the claw.md. - You can actually take everything that's in your - claw.

md - and put it in a skill.md - inside Cowork. And you go, oh, well, that's - not helpful. Like, you're all you're doing is - telling me that I'm just using co work - to process my clippings and things like that.

- I actually recommend going a step further. Create - yourself skills - that do things like - take take the emails in my inbox - and - save them into my ingest directory as markdown, - and it'll take all your emails for that - day. Like, I've I've somebody on my team - did this, and it's super cool. He has - it so Cowork - looks at his inbox every day.

It takes - all the items in his inbox, - and it turns them into markdown. And then - once they've been turned into markdown, they just - get marked as read and archived because now - they're sitting over in markdown, right, just in - in his OneDrive. - And then he runs an LLM Wiki style - kinda thing on top of it to - extract what are his priorities for the day. - Maybe he got an email from me or - he got an email from a coworker or - he's got specific projects that he flagged, things - like that.

And then he has this set - of skills that iterate over it, and I've - been playing with this too. It actually works - really darn well. - You can take those things, - reason over them, create your own personal knowledge - management vault just inside of your OneDrive for - business. Obsidian's free.

You're effectively getting cloud sync - because you got markdowns. So if you're in, - like, you're in, like, oh, hey. I need - to get to this on my work phone - or things like that, run Obsidian on your - work phone. Just point it at your One - OneDrive directory.

It's okay. - It it'll all work. And then you get - to start thinking about, like, richer functionality within - Obsidian. So this was the thing that unlocked - it for me in the Copilot Cowork scenario - was not just, like, reason over what happened - today and maybe give me a, like, a - table.

Like, tell me, like, the top three - people who emailed me or things like that. - But Obsidian has plugins - and - Wiki style markdown for things like task lists. - So you can actually have it, like, take - if you got 50 emails last night and - you walk in the morning, - you can have it take those 50 emails, - reason over them, do whatever it needs to - do, and create, like, your to do list - for the day. And then you actually get, - like, a clickable to do list, not in - Planner or anything like that.

Like, you can - just put it in Obsidian and have it - there and ready to go, and it's very, - very portable. And because Copilot Cowork - has access to, like, your team's messages and - your inbox and all these things, You can - have it, like, mark your email as read - and archive it. You can have it organize - your email for it as it's doing this - extraction and pushing things out. You might have - it like, in my case, I get a - lot of, like, junk mail and newsletters.

- So no matter how hard I try and - filter those things out, some always make it - into my inbox or my inbox other things - like that. So I can have that flow - come back and - clean that up as well. Right? Like, say, - hey.

It notices there's a newsletter there. I - don't have to worry about creating a custom - rule for it anymore. I've just got Cowork - dumping that stuff off to a newsletter directory. - And then part of my ingest flow, because - it's still ingesting those things into my Wiki, - is telling me what newsletters I got that - day just in case there was something there - that I wanted to read or or things - like that.

So you can absolutely take this - concept - and pull it over to - the world of things like Copilot, Cowork, - and all of that. Or if you wanted - to yeah. I've seen a couple folks do - this with - just something like GitHub Copilot and WorkIQ - or - Copilot - CLI, - Claude with the, like, the WorkIQ MCP and - things like that as well. The problem with - the WorkIQ MCP is I've had way better - luck with, like, Copilot coworkers saying, actually, like, - move this message from here to here.

WorkIQ - is, like, glorified graph reader. So you end - up in some weird throttling scenarios and other - things that CoPilot CoWorks seems to generally skirt - around. - You're out of time. CoPilot needs to be - deployed ASAP, but is your tenant really ready - for it?

Years of data, permissions, and users - lurk in its shadows, ready to be exposed - by AI. - Sharegate Protect sees it all, so you can - find the exposure risks, fix them fast, and - deploy AI with confidence. - Microsoft - three sixty five governance. We got this.

Learn - more at sharegate.com/governance. - 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 - Yeah. And then just use - so you use Copilot Cowork for everything.

Yes. - Use Copilot Cowork as the harness. As the - LLM, the harness. Yep.

Where I said, hey. - When we started in the beginning, I said, - hey. I was doing this in Claude with - open router. Replace Claude and open router with - CoPilot CoWork, - the Frontier agent.

And then you just switch - from doing everything in the CLI to going - into Cowork and starting - conversations - in Cowork in the browser for all the - ingesting and - processing that raw data. Yes to starting. But - once you're like - let's do, like, the crawl, walk, run thing. - So I think crawl is us walking in - there and typing all the time and kicking - these things off.

Once you have a refined - workflow, Copilot Cowork lets you schedule this stuff. - And if you're structuring it in a way - where you're being principled, like in my case, - with my local - vault here that I was using for, like, - ingesting RSS feed articles and things like that. - So in that one, - I'm sitting in, - I'm sitting in a world of that workflow. - That's really like an on demand thing.

Right? - Like, I don't always have clippings in there. - I don't need it to be looking once - an hour or anything like that. Emails are - coming in all day.

So I can have - that skill - run every hour and do that ingest - hourly. So I don't even have to live - in Outlook. I can just go live in - Obsidian. It's kinda kinda kinda nice.

The other - thing too is - I talked about the concept of, like, taking - some of the heavier operations and going and - putting those in scripts and things like that, - you can still do that in co work - because co work can execute local flows. - Or the other approach that you could take - would be - to really think about maybe rather than, like, - scripts and functions - to just skills and functions. - So having kind of the building blocks sitting - inside of composable skills - that then you can stack per scheduled task - or whatever you wanna do on on on - the CoWork side.

Interesting. There was another question - I was gonna ask about this, and I - can't remember what it was. Adjusting, scheduling, - oh, - have you compared - so this has always been a debate too, - right, is how - Copilot does in terms of responses versus something - like Claude or - any of those. 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 I know CoWork is different because a - lot of it I don't know that there's - as many system prompts injected into - CoWork - given that things like DLP don't even work - in it versus the normal Copilot.

I feel - like the system prompts are are still there. - You do have - depending on your tenant configuration, - you do have model selection or at least - model I would call it model selection light. - Like, in my case, I can choose a - couple variants of GPT - and a couple - anthropic models. - You do get much richer model selection when - you're over on a CLI - or, like, a GitHub Copilot chat kind of - experience.

- But the nice thing is - you're talking one vault that just lives out - or a series of vaults. It's all just - text, right, that that lives out here in - a side card. So you're not restricted - to - just Copilot co work or just GitHub Copilot - CLI. You could augment this.

Yeah. You can - augment this and kind of go in. So, - like, for me, I tend to live in - the terminal. - So I'm not going in, like, during the - day.

I don't actually, like, worry about going - into cowork and to ever typing a question - in a cowork. Cowork's a harness for me - to go and run skills, do the extraction, - kinda Right. Compound value and compound knowledge over - the course of a day, a week, a - month, things like that. - But when I'm going in and I'm going - across - and I'm saying, hey.

Look at this stuff. - I'm typically doing that through - the CLI or, like, just just, like, GitHub - Copilot chat, things like that. Now the other - interesting thing that I found here is - it starts to make the data that you - extract - a little bit more deterministic. - So - it's a weird black box when you go - into Copilot - or you go into co Copilot cowork, and - you say, take a look at all my - emails from today - and - extract StackRank my top three priorities.

Right? You - never know. Is it actually gonna search all - my emails for the day? What is what - does it consider an email for a day?

- Is it only looking at my inbox? Is - it looking at my inbox plus my inbox - other? In my case, I run inbox inbox - other, and I do the Scott Hanselman inbox - cc things. I got I really got three - inboxes.

Right? So I never know, like, what - it's looking at and what it's doing. But - by extracting the data and pulling it out - and sidecarring it, when I say go look - at today, today is a very real thing. - It's a discrete set of data that sits - out there and is available.

- And and I found that accuracy has kind - of improved there. I have, like, far more - confidence that it - can can do what it needs to do. - Have you run into - anything, - like, using co work to your point earlier - about it can be expensive? Sorry.

I think - I have a meeting up. - This episode is brought to you by Trusted - Tech. If you're managing Microsoft three sixty five, - you already know licensing isn't simple. - E three versus e five, the brand new - e seven licenses, - Copilot security bundles, and with price changes coming - in July - 2026, the stakes are getting higher.

Trusted Tech - helps IT leaders make sense of their Microsoft - three sixty five environment with a free licensing - consultation. - Their engineers analyze what you're licensed for and - what you're actually using and where you can - optimize, whether that means consolidating - security tools, preparing for Copilot, or eliminating wasted - spend. Plus, they offer proactive - and reactive support, which has been awarded yet - another solution partner designation from Microsoft - for support services.

- If you want a clear, data backed plan - that has helped over 7,500 - subscribing - customers - save up to 20% on Microsoft three sixty - five before your next renewal, - visit trustedtech.team/mscloudprom365 - and schedule your free consultation - today. - Okay. I have a meeting at four.

Where - you run into some of that quotas and - stuff? Like, 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. - It was crazy how quick I burned through - tokens. - Did you - how much did you test cohort to see - if during the ingest or have you noticed - do you start getting - throttled or limited where it's like, okay. Go - use something else to your point for that - initial if you're gonna ingest a bunch of - stuff, or do you have to start being - careful about how often or how much you're - ingesting yet?

- I can imagine it might you might start - getting limits eventually. Yes. You do have to - be careful. So so I think the - interesting thing here and the callout that I - keep coming back to is - and you're nailing it, is the context windows - and the token usage, both on inputs and - output, does get very heavy.

- So - I I think this is someplace - where I would be inclined to think that - folks in our community would be able to - wrangle this pretty well given kind of their - backgrounds as - developers, - IT pros, more experienced end users. You have - to think a little bit like that scriptwriter - or that programmer. - So your job isn't just - to have a big context window that you're - firing off every single time. It's kinda like - ruthless optimization.

- Get this down to here. Get this down - to here. Get this down to here. So - so what I realized was - when I went and did this on my - own, just kinda same thing over in, like, - Claude with open router, blah blah blah, and - I too had to switch over to my - paid Claude plan.

Don't get me wrong. Like, - it it it didn't get me all the - way there. It's expensive. - I think one of the reasons it didn't - get me all there all the way there - because I didn't think about, like, a programmer - from the start.

- And I I just kind of approached as, - like, the spirit of what mister Kabathea put - out there of, hey. Here's an LL on - Wiki, and you give it a simple set - of instructions, - which I I recognize a lot of these - folks who are putting these things out there. - They do live in a world of unlimited - tokens. Right?

They work at Anthropic, and they - work at OpenAI - and things like that, and not all of - us have that luxury. - So you do have to kinda think about - that and start to refine things. I have - noticed - I I will say so I I mentioned - I've built a couple of these now. - So the first one, I really went down - that refine refine path.

Now I have this - really solid - templated thing that already has the scripts from - day zero. So when it starts, - it's already starting in a more optimized state - rather than having to start from scratch every - single time. So you kinda carry that knowledge - forward, and that helped me a whole bunch - when I was thinking about the Copilot Cowork - side of it. And then it really started - to click for me when I saw one - of my coworkers - who was in the mixed modality of Copilot - Cowork - for things like email extraction, - but then they were on they were still - on the CLI and doing, like, native Obsidian - Web Clipper things and all that stuff.

Right? - Like, they were kinda using multiple tools - all to have the data coalesced - into - a single source. - And, effectively, like - I mean, Obsidian's markdown, you can make it - look pretty, all those things. It has checkboxes - and tables, and you can put images in.

- Like, is it as rich text as something - like OneNote is? Absolutely not. But when you - start to get this flexibility of having, like, - multiple ingest points, multiple output points, - and kind of the capabilities that come with - the plugins in Obsidian, - all for free, - it kinda just works. Right?

It's a very - it's it's a very cool system. - And once you start to go down this - path, like, for me, I, like, I'm one - of those people, like, I hate to admit - it, but I can live in email. Right? - It's pretty easy to get 50 emails overnight, - then you just get stuck in it.

You're - like, I don't really live in email as - much anymore. - Email has just turned into I've got this - other thing over here that's taking a look - at it. I had to refine it. I - had to take some time put time into - automating that process and - automating myself out of that job of living - in email, - but it - it helps on the other side.

And I - think, like like, the thing that really, like, - kicked my butt was - this team teammate of mine who, - he lives in the chaos that I live - in - and the rest of our team lives in. - And he has this - super awesome - I hesitate to call it a demo because - he showed his real workflow for how he - manages all this stuff. He put it all - in a Git repo for us so we - could just pick up skills - and run with them. And he's one of - the first people that I'd seen who successfully, - like, - navigated - inbox zero - at Microsoft - without feeling like he was missing out on - something or feeling like you're not just doing, - like, the - control a, delete, or control a, archive, and - I'll hope they email me again tomorrow - kind of thing.

- It it was also very interesting to me - to see the way that - this individual managed tasks in Obsidian, like, how - they weren't trying, like, a forcing function of, - like I've seen a lot of people use, - like, Copilot cowork and and integrate it with - Planner and things like that. They put together - their Kanban boards, and I I feel like - the process runs away from you. It gets - a little too fancy at some point. I - I think - we're much better served when stuff becomes nice, - simple.

Like, here's a to do list. I've - got a checkbox. I check that thing off, - and it gets a strike through it. - And you get to feel good about that.

- Like like, there's a little there's a little - rush of endorphins, - right, when you when you check that checkbox - and have that - sense of accomplishment that's out there. And because - this stuff is all text, you can do - things because, again, LLM Wiki style Obsidian, you - don't even have to check the checkbox. - I would imagine I haven't done this, but - I think this would be totally doable. Like, - I could go into GitHub Copilot CLI, and - I could say, - my my task for today - to email Jack is done.

Can you just - check that off for me? And I would - go check it off for me. Right? I - don't have to - worry about actually even opening a vault or - looking at anything and - running that way.

- Like, the the elephant in the room, which - I don't think anybody could address at this - point, - and, I mean, it it is there, is - your like, this stuff's all plain text. Right? - You are extracting it out and and kinda - putting it down to your, - your OneDrive and things like that. So stuff - like - IRM and MIP - really don't apply to markdown files.

- So it certainly, like, don't break, - like, organizational constraints - or things like that. Pay attention to sensitivity - labels and all that stuff. Those types of - things? Well, Cowork doesn't pay attention to sensitivity - labels, so why should I?

No. It does. - It can't read any MIP file. It's kinda - useless.

So Yeah. If you're actually encrypting it. - Yeah. No.

I'm gonna have to play with - this too because I'm doing all my meeting - I'm doing this with meeting notes, and I - created a Wiki for meeting notes. So, like, - meeting notes out of people, so I have, - like, people and topic and all of that - meeting notes. It would be really interesting to - take this into Cowork and actually, like, now - I can correlate meeting notes with emails, with - all of that. How about just meeting notes, - but meeting notes with transcripts?

So go create - yourself a skill, load it into co work, - have it extract your meetings transcripts - from the prior day, ingest them in, and - then go run this over them. Right? Like - like, that's just a super Yeah. I see.

- I'm already doing that part of it. Now - we just have to tie in the emails. - So it is. This it's it is definitely - interesting to start tying all this together.

- Again, those are the two I've done is - emails. I did the podcast. - I've started to do just references, but I'm - gonna go play with blog posts and figure - out how to incorporate some of that other - stuff in with Cowork. But I just gotta - chat that my other meetings started.

World of - possibilities here. Absolutely. So I'll put a bunch - of links in the show notes to kinda - the original gist and things that came out - here. Yeah.

There's some good YouTube videos out - there we can post in the show notes - too, - walking through how different people have done this - and showing how some of it's set up - too. I haven't seen a lot of people - who have done - Copilot Cowork. There's a lot of Claude. There's - a lot of good reference implementations.

- Yeah. Claude's out there. Super easy to, like, - Google, Bing, DuckDuckGo, Kagi, whatever your search engine - choices. - I will put out some examples for those - who are in the Microsoft ecosystem.

- So So there's a pretty good one out - here that I'll put in. It's called the - GitHub Copilot Wiki, an AI powered second brain - template. - So it that's it's kinda like a forkable - template that you can start with. So it - already has the wiki structure.

- It has a template dot m d. - It has a lot of these constructs, right, - that I talked about here, and it can - just kinda, like, bootstrap you if you didn't - want to - if you didn't wanna start start for yourself - links a template too. So we'll drop those - in the show notes. Cool.

Perfect. Well, thanks, - Scott. Yeah. Enjoy your weekend, and I'm gonna - go jump on my next meeting.

Perfect. Me - too. - Life life is on. Alright.

Thanks, Ben. Alright. - We'll talk to you next time. Yep.

Bye - bye, Scott. - If you enjoyed the podcast, go leave us - a five star rating in iTunes. It helps - to get the word out so more IT - pros can learn about Office three sixty five - and Azure. - If you have any questions you want us - to address on the show or feedback about - the show, feel free to reach out via - our website, Twitter, or Facebook.

- Thanks again for listening, and have a great - day.