What I’m thinking about: Google Drive swore my Claude skills were synced while Claude quietly ran old copies of a dozen of them…and one line pasted into Claude Code now sets up the fix for anyone who wants it.
Claude is probably running an old copy of your skill (on September 30, almost 20% of my skills were outdated…and I thought I had good skill hygiene).
If you’d rather watch than read, here’s the full demo (or click the image below) on how to install one of the most useful builds we’ve ever done, start to finish on a fresh Claude account. Both jump straight to the install at 9:57, and the install itself runs about 8 minutes of video (a couple stumbles included).
https://www.youtube.com/watch?v=iWvV-8Iyaj8&t=597s
Last week I mentioned that my Claude Code setup auto-calls skills (saved playbooks for specific jobs) from a GitHub repo. This is that repo, plus a free template that builds your own from one pasted line.
How the drift snuck up on me
Most people building with Claude end up with a growing pile of skills in a Drive folder, in Notion or uploaded as .skill files, with no clear sense of which copy is the most up-to-date version (add team members with different permissions, and it gets messy FAST). Before building this solution, I’d have included myself in that camp too.
A few weeks ago, I compared every copy of every skill I had, and the skill I’d edited often wasn’t the skill that ran. Twelve had edits that never reached claude.ai (the Chat/Cowork version of Claude), and Claude Code was still loading weeks-old copies of the validation checks that fire on every build decision…so Claude Code, Claude chat and Cowork weren’t even running the same versions.
Google Drive, meanwhile, showed every skill file as synced the whole time…gaslighting me into believing my skills were current.
One repo, every surface, no hands
The fix is one private GitHub repo (a project folder that saves every past version of every file) set up as a plugin marketplace, a catalog that Claude Code, claude.ai and Cowork all subscribe to, like following a feed. All 69 of my skills and 21 scheduled routines live there now, and small scripts called hooks raise the version number and send each change up to GitHub, so an edit reaches every surface without a single upload.
I built it in an afternoon after Kristen Parise, a colleague of mine from The Uncommon Business, sent me Matt Pocock’s skills repo (highly recommended for making your agents and CLAUDE.md files more accurate and cost-efficient), and I riffed on what his repo contained.
Setting up the skill auto-updater in Claude is less user-friendly than you’d think, so I turned it into a template that strips out as much of the friction as possible (backed by 10+ years directing engineering teams, plus years of living inside Claude).
I update skills and routines every day, sometimes tens of times a day. When I tell Claude I didn’t like how something was phrased, it updates my voice skill, and the change reaches every surface on its own…no downloading, no re-uploading, no wondering whether Drive saved it. It all just works, and I can’t tell you how much brain damage it’s saved me.
Set it up with one line
1. Count where your skills live. Google Drive, Notion, OneDrive, the uploads on your Claude account…more than one home means drift, and if it hasn’t happened yet, it almost assuredly will.
As the old adage known as Segal’s law goes, “A man with a watch knows what time it is. A man with two watches is never sure.”
2. Paste one line into Claude Code. This one:
Set up my Claude skills repo from github.com/serious-ai/claude-skills-starter. Read its SETUP.md first and follow it step by step.
Budget 10 to 15 minutes for the install itself (longer if you’re bringing in a big pile of existing skills).
(Never opened Claude Code? In the Claude desktop app, it’s the Code tab, and it comes with Claude’s paid plans. If you don’t have a GitHub account yet, sign up free at github.com before you paste the line.)
Claude checks your tools (and walks you through signing in to GitHub), then asks three questions, each with a suggested answer you can accept as is:
– Repo name (your skills’ home on GitHub), e.g., claude-skills
– Private or public (I’d go private, since skills tend to hold client names and house methods)
– Plugin name (just the label your skills show up under inside Claude), e.g., your company name, like acme
Then it makes your own copy, brings in your existing skills and saves them to GitHub.
(In my video walkthrough, I ran it in Terminal only because it was signed into a separate Claude demo account. Plenty of people find Terminal intimidating, and I’m not routinely in it myself…the desktop app is easier to follow.)
Btw, if some of your skills only live on claude.ai, Claude will ask you to download each one or paste its text in, so have those handy.
3. Run the two commands it hands you. Claude Code’s safety rules won’t let it install plugins into itself, so it hands you two short commands: the first tells Claude Code where your skills live, and the second installs them.
Paste them into Terminal, or click run on each one in the Claude desktop app. Then type “/plugin” in Claude Code, open Marketplaces, pick yours and turn on auto-update, so every future edit arrives on its own. Start a fresh Claude Code session afterward, and your skills will be there (it’s copy-paste, and Claude corrects itself if a line doesn’t take).
(The CLAUDE_CONFIG_DIR part in my screenshot only kept the demo separate from my main setup. Yours will be two shorter lines.)
4. Add it on claude.ai (for chat and Cowork). This is the one step Claude can’t click for you, so here’s exactly where to go:
Customize → Plugins → + Add → Add marketplace
In the URL box, type your GitHub username, a slash and your repo name, e.g., janesmith/claude-skills (the repo name from step 2, not the plugin name). Leave “Sync automatically” on, then click Sync.
Not sure what your username is? Ask Claude Code, since it signed you in during step 2.
If your copy is private, connect your own GitHub when asked, limited to that one repo (a public copy needs no connection, and you never connect to our template). In the desktop app, do this from a Chat window, since the Code tab’s Customize screen is a different place.
Then open the Discover tab, find your plugin by the name you picked in step 2, and click ‘Add’ on its card (that’s the actual install). Your skills show up under it (FYI, on a Team plan without admin rights, this screen may be blocked, so ask your admin).
Then go back to Claude Code and tell it you’re done, so it can finish up.
5. Install first, delete second. Once the plugin lists your skills, delete your old uploads on claude.ai (or from a Chat window in the desktop app) under Customize → Skills → “Created by you,” so nothing runs twice.
That’s the only cleanup you do by hand. Claude moves the old copies on your computer, the ones the desktop app’s Code tab shows, into a backup folder at the end of setup (my demo skipped this, since a fresh account had nothing to clean up). From then on, you NEVER upload or install a skill file again…everything lives in GitHub.
To confirm it all worked, open a new chat and ask Claude to list the skills it has. Yours should be there.
6. Add new skills through the repo. When someone sends you a skill (a link, a GitHub page or a file sitting in your Downloads folder), hand it to Claude Code instead of clicking install. Paste the link, or drag the file into Claude Code, along with this line:
Add this skill to my repo, and read every file before it goes in: [link or file]
If the skill lives in a shared Google Drive (or other cloud storage) folder, download it first, since Claude Code often can’t open a Drive link.
Read it before it goes in
Serious AI Partners’ motto, ‘Underwriting Over Everything,’ also applies to your tools…nothing ships that nobody examined and nobody owns, including the skills running your business.
A skill is instructions (and sometimes scripts) running with your agent’s access, so a downloaded one can carry a prompt injection, hidden instructions telling the model to do something you never asked for. In a February scan of 3,984 public skills, the security firm Snyk found 13.4% carried at least one critical security issue, and 76 held confirmed malicious payloads, built to steal credentials or open backdoors.
Even good actors get hacked, and no review catches everything. In June, researchers at Trail of Bits slipped malicious test skills past the scanners from ClawHub, Cisco and skills.sh, and concluded that “No amount of scanning or LLM analysis can reliably detect malicious content in agent skills.”
So a read-through won’t be 100% reliable…but one door in, every file read and the source written down beats blindly installing random skill files from folks you may or may not know (where the odds of picking up malware run far higher) by a WIDE margin.
Two traps, in plain English
I hit every trap along the way on the initial build, and the template carries a fix for each. A couple of examples:
Adding isn’t installing. Adding the marketplace installs nothing, and your plugins wait under Discover until you click ‘Add’ on the plugin itself (on the demo, I briefly wondered whether I’d installed it at all…this is one of the less user-friendly areas).
The pull request trap. claude.ai only updates when a change arrives through a pull request (GitHub’s “propose a change, then accept it” step). Eight of my saves sat unseen for two days before I caught it, so the template opens and accepts its own pull request on every save.
That base (a setup guide written for Claude to follow, plus hooks that handle versions and syncing) is why one pasted line can carry the whole install. Mid-demo, Claude even logged seven friction points a first-timer would hit, which I took straight back to my main account (fix all of this and make sure there aren’t any of these issues anymore)…and all seven were merged into the template the same day. That’s my constant process for any issue that shows up in Claude.
A rule becomes a check
One idea I took from Matt’s repo is that a mechanical rule shouldn’t live in a prompt. I used to tell Claude in plain prose never to use em dashes, which held most of the time (and cost tokens on every run).
Now Vale, a free style checker (think spell-check for your house rules), holds that rule in a nine-line file, and any email draft with an em dash gets stopped before it sends. A yes-or-no check holds EVERY time.
Down to one watch
The template also lays the groundwork for companies, with one shared repo, a plugin per department (marketing’s skills for marketing, HR’s for HR), and an owner who approves every change before it reaches anyone.
On a Claude Team or Enterprise plan, an admin can sync that repo into the whole company and set a plugin to install for everyone automatically (or, on Enterprise, for just one department), so nobody uploads a skill by hand. The split matters more as a library grows: past a certain size, Claude Code starts dropping the descriptions of the skills people use least, and Claude stops matching those skills to requests.
…Speaking of companies, Serious AI Partners hit a milestone last week, as payment landed from our first implementation client, about a month after this business launched. It’s the fastest I’ve gone from a new business idea to real money in (and not pennies, either), of anything I’ve started…and money in the door is the best proof of concept there is (particularly since I run such a lean shop, and we’re already cash-flow positive).
As a final reminder, the first step of your own Claude skill-updater install is to paste this line into Claude Code (budget 10 to 15 minutes):
Set up my Claude skills repo from github.com/serious-ai/claude-skills-starter. Read its SETUP.md first and follow it step by step.
=====
If you run a ~$5M+ business and want game-changing tools like this (e.g., one source for every skill, an owner approving every change, and checks that hold on every run) implemented across your teams, that’s the work we do at seriousai.partners. It’s how you become the operator whose AI runs on the current version of your judgment, every time.
Regardless, the template is free. Paste the one line, and Claude will walk you through the rest, the same way it walked me through it.
(If the template saves you time, click the ☆ Star button at the top right of the repo page. It’s free, takes one click, and helps other operators find it:)