Serious AI with Chris Duff
October 14, 2026
Custom Claude skills tend to live in several places at once: a Google Drive folder, a .skill file someone uploaded, and whatever copy Claude actually loads. I had around 70 skills and 21 scheduled routines spread that way, and Google Drive showed every file synced while Claude Code ran different versions than Claude chat and Cowork. This episode is a screen-share of the fix: one GitHub repo that every Claude surface installs from, set up from zero on a fresh account with the free claude-skills-starter template. I'm Chris Duff, an AI builder with a medical background who has led engineering teams since 2015, and Serious AI Partners runs on these skills every day.
Prefer audio? Listen to Serious AI with Chris Duff on Spotify, or wherever you get your podcasts.
(Podcast transcript below)
Claude is probably running an old copy of your skill or routine. Welcome to Serious AI with Chris Duff. Today I'm going to be going over one of the most important builds I've ever done that just makes working through updates on Claude and, you know, really any type of AI setup, far more effective and efficient since you're always working on your most updated versions of your routines and skills and not having to worry about going through manual updates or making sure there's other routines, syncing skills and across different data storage silos or anything like that, whether it's via Google Drive or Notion, for instance.
So I want to walk this through with you all and do a full-on demo here. So, again, I think anybody who's routinely using Claude or ChatGPT, it also has skills, is that you'll have a growing pile of skills, usually kept within a drive folder, or they're just uploaded as .skill files, but then you're not sure where that original lever is kept or when to update it, especially if you have team members and different permission sets on who should be able to access skills and so forth. Like it can become a huge mess very quickly here. And so a lot of folks, like as they're moving along in their day-to-day, it's just easy to get things lost.
I would include myself there for, you know, really since skills had come out on Claude. And so my role here is an operator where I had around 70 skills that were in relatively routine usage. And I also had 21 separate routines that were running on my Claude Code set up. And I was able to move all of those into a GitHub repo. So, everything was able to automatically update without needing me to approve anything or do any extra clicks or anything like that, it was just all moving smoothly in the background for me. I hit every trap along the way to do this, but I'm giving away my template on how you can implement this precisely within your business, because again, I think this is one of the most useful things.
I've built so far here, and not as user friendly as you might think to do this, but I try to remove as much of the friction points as possible. And so, again, like why this really matters here as well too is, you know, for much of the time that skills were available, many times when I might have edited a skill, It might not have been the one that was actually running sometimes like Claude Code would run different versions of the skills compared if you're in regular Claude chat or Cowork for instance It wasn't one-to-one But my Drive session was still showing that you know Google Drive was still showing that all my skills were Synced the whole time so it was it was just confusing on why everything wasn't necessarily updated.
And again, GitHub is able to, you know, solve this issue here and allow my libraries to keep growing and make sure that across any Claude instance, chat, Cowork, well, chat and Cowork are the same thing now, and Claude Code all have the most updated version of the skill and routines all work within Claude Code. And again, I have 10 plus years of experience leading engineering teams in my business. Serious AI Partners runs on these skills every single day and routines. Routines include a whole bunch of skills. And I built all of this within an afternoon a couple of weeks ago.
And shout out to Kristen Parise, a colleague of mine from the Uncommon Business who sent me Matt Pocock's skills repo within GitHub and this really inspired me to create that repo that I just mentioned that auto updates skills and routines automatically here and I would highly recommend Matt's repo here too, it's very efficient for helping your agents and CLAUDE.md files to be more effective and cost-efficient in the way that they operate. Really bumping up accuracy here, like how agents should actually talk to each other. So that was kind of the whole point here, and it just led me in kind of a different direction once I really started playing around with it.
So besides just, you know, you guys get like the whole purpose on what I'm doing here. So here is the public template that you're, you know, you can access at any time here, which I will include in the show notes. This GitHub claude-skills-starter from my Serious AI author profile here. And so you, you have the whole read me on what exactly happens here. And so I, if you're interested in seeing, okay, how you'd actually go ahead using it. Like this tells you exactly what to do. But again, I wanted to make this as simple as possible and having AI handle the most technical aspects of installation is generally always going to be the best case here.
So now I have a one line install prompt that will help you get this set up on your machine and it'll prompt you accordingly for pieces that might need your permissions or direct clicks and so forth. Another key piece that I just want to note that I hadn't really thought about before is that once you have all of your skills and routines running from GitHub instead of locally on your Claude account is that there will be an automatic check to see whenever you install a potential skill file because so many of us like that we might just be downloading from random people here and unfortunately there could be potentially malicious prompts, prompt injections that could lead to nefarious actions that you wouldn't want happening on your data here.
And so if you're working through a repository like this, it will first assess any of those skill files and determine okay, is there anything malicious that should be removed from the underlying skill before you just blindly install that and I'm glad nothing actually had happened when I have installed a whole bunch of skills from folks who I think we're good actors here too, but sometimes some of these larger good actors might be getting hacked themselves facing cybersecurity risk and then a malicious actor can get into their file. And so it can spread a whole bunch of potentially downstream issues.
So this is another benefit just to make your cyber environment even more secure here. So I'm gonna go ahead and follow my own advice here as far as, okay, just set up this Claude skills repo. And you can, normally I'm using the Claude desktop app and because I have some certain routines working in the background, I have a separate Claude account, but I'm not gonna be using the primary desktop app. Usually I'm using Claude Code within it. Like you can see here, like I'm in the Claude Code set up. So usually that's my game plan here. So I am going to be using terminal instead, which is already logged into my other Claude Code account.
So a lot of folks like get intimidated by terminal here. I'm not routinely in this. Some aspects of Claude Code can only run through Terminal and it gives you even more firepower for some just deep engineering tasks that the desktop app or the web app, certainly you can't even run Claude Code on the web app, wouldn't necessarily allow. So, but you can see even in terminal, I can just go ahead and, and I'm actually switching over to my other profile here. So this is my other Claude account that is logged in, pretty much a fresh account. And so, you know, that this terminal is already connected to this account here and so it's just going to prompt me whenever it needs me and again this is like totally fresh I don't even think I have a github account for this email so this will basically just show you precisely what you would need to do to install this thing fresh and so I can just show you okay terminal is just going to be running here.
I'm actually going to pause just to keep this concise until it actually needs me. All this is being run automatically right now. So you can see it asked me a couple of questions here for decision. I'm just answering that back to make sure that we're proceeding with a demo here that's still gonna be going through the entire process and then getting rid of it. So this is another example on how you can, you know, if you ever want to do demos within Claude, hey, you can already arrange for getting rid of unnecessary files once you're already done with it.
It's really effective to do presentations like this. Okay, so went through a series of steps here. Again, I did not have this paused for very long at all. It's asking me to throw a couple of these items into terminal. Again, sometimes the desktop app will throw you certain commands that you can run within the desktop app. That's just inherent to Claude Code if you want to be able to run certain items like this, but it's copy-paste. It's still pretty simple to do. So then it's asking me. Okay. Okay, so it's asking me to fix that. Do this. See, I mean, Claude Code, like it can auto-correct itself as well, too, if it doesn't have the right type of line that it gives you.
Okay, so then it's asking me to do a couple other things within Terminal, including plug-in, So, yeah, you can just do those slash commands and you'll see it'll pop up there and you plug in. What was it asking me to do? Yeah, then I can click on, or no, yeah, Marketplace, scroll down, and I'm just using my arrow keys, for instance, on how to do this. So I wanna enable auto update. So that's all I need to do. There. And then I can get out of that. Just kind of like escaping or going back effectively and then it's thinking that I'm the desktop app so I need to tell it that I'm not.
I'm on claude.ai in the browser signed into the TUB account, not on the desktop app. So, now I'm going to move this over just to follow the directions that it was saying. I'm going to customize. This was the part that was certainly tricky. Some of you might be wondering, like again, I'm trying to make sure that it is as streamlined as possible following. Again, even if you're looking at this terminal, like don't be as alarmed because the UI is easier to follow on the Claude desktop app for instance. So we're on this, then it says click on plugins and then I'm going to go to add marketplace and if you're on a teams account and you're not an admin you might have this blocked so just FYI for that if you are then you might want to ask your admin about this.
This mainly affects the skills being auto updated for Claude chat and Cowork instead of Claude Code to me, like I'm in Claude Code all the time, so this would be less important, but just FYI on why we would be doing this precisely. And so then I want to add in serious-ai/claude-skills-demo and then I will sync it. Okay, if it's private, so it actually is asking me to connect my GitHub. So in this case, okay, so now I have GitHub connected here. So I'm going to sync that. And if you have your own GitHub account, you will need to potentially connect that.
It's generally recommend it. GitHub accounts are free, so you can always set one up. So you can see that Marketplace, well actually it hasn't been added here. So then you also have to go through and add the, so then you go to yours, the customize, okay for Marketplaces, it is fully added now. So see like it's a little tricky in terms of making sure you're clicking on everything correctly. But then I need to go back, tell it, okay, that's added. And now I need to install it. So yeah, this was the other piece here. So I have to go to, okay, discover.
Did I not install this? So now I feel like I'm a little gaslit. Ok, everything has been installed here. So both skills are listed. So yeah, and then it's getting back to its full review. Okay. And the demo is installed with both skills here. See, even running through this, it was noting a couple of friction points on to make this whole demo process even smoother when you tried doing this and this all be fixed. So then I would take this back to my main Claude account. Okay, fix all of this and make sure there aren't any of these issues anymore. So that's like my constant process on fixing anything that might show up in Claude.
So now with all of that done, that's actually everything. So that was actually a much simpler route than when I was first building all of this out. And so I'm not gonna necessarily go through the remainder of the demo here but anytime let's say, you know, like this build checker house voice or whatever. Anytime I might have, you know, a back and forth with Claude of like, yeah, I didn't really like how that was phrased. It will auto update the house voice, go back to the GitHub repo and make sure that it's fully updated. You don't have to do anything so that it'll never have to prompt you, Hey, you to download the skill file, re-upload here, and making sure that it's synced across your team, or again, across Claude Code and regular chat.
It all just works. Same thing with the routine. So like I am updating my skills and routines every single day, sometimes several times a day. Like sometimes tens of times a day because of just how frequently I might be validating or again, adjusting and calibrating. So it used to be super annoying to have to and I would get so much skill drift Just like yeah, did I actually update this is this properly saved and drive and all of that? It's just so easy to get confused. So this solves all of that and Additionally that there were some of those items from that Pocock's repo That were installed Such as again having the Vale style checker.
So for instance, like I used to always have like, yeah, don't use em dashes and so forth. So now the, with Vale, it just sets up a automatic check on anything going out. So it doesn't have to rely off of prose rules that might not be a hundred percent accurate every single time and waste tokens. Now it just checks, you know, are there em dashes? Yes or no. With a final check before anything goes out, and that will go across any pieces of your content that you produce within Claude. So these are some other useful additions that happen here. And also within that public GitHub repo, I also included an option for Teams.
So this was a bit outside the scope for today here, that can be a separate session, but there is the groundwork laid for, if you're managing a team, how do you make sure that different plugins are installed across particular teams? Maybe you only want permissions for certain GitHub repos to go to your marketing team versus your HR team, whatever. So that's another avenue where you could use like this same system, customized to handle different permission sets here. So hope this one was helpful here today. Again, I can't tell you how much brain damage it saved me over the past couple of weeks since I've installed this.
So for you, like the next step here is like count all the places your skills live. Do they live in Notion? Do they live in Google Drive? Do they live in SharePoint? For instance, are they just on your Claude account and you just install them and delete the skill file and who knows where they go. If you have more than one place where they're stored that means drift inevitably will happen. If it hasn't already, almost assuredly it has. And so again, you can just open up Claude Code and paste that one line that's in the repo and just set it up exactly how I demoed it for you.
And Claude will walk you through the rest the same way it walked me through it just now. So I hope this one was helpful. Give me a thumbs up or reach out or any potential questions how I could make this smoother. Subscribe and share everybody. And again, Serious AI Partners for enterprise builds for mid-market companies, 5 million and up run rates. So would love to work with more of you. Take care all, bye.
Fill out a short form, and we read every answer ourselves. If there is a fit we will reach out and find a time that works for you, and if there is not, we will say so and point you toward someone better suited. Either way, the moat question is worth answering now, in a market AI has already reset.