Skillsmaxxing: How Remy Runs His Whole Company on Claude Code Skills
Greg Isenberg has a word for the biggest AI unlock of 2026: skillsmaxxing.
And no, it's not a new model or a smarter prompt. It's skills. The boring markdown files that tell your agent exactly how you want things done.
His guest Remy (you might know him as AI with Remy) runs his entire company on them. He lost 150 of them in one afternoon, rebuilt the whole system in GitHub, and now distributes every skill to his whole team through a single plugin.
I watched the full episode so you don't have to. Although honestly, you should.
Skills are just SOPs for AI
Remy's definition is the cleanest one I've heard: a skill is an SOP for your agent.
The old way looks like this. You ask Claude to write a sponsorship proposal. The first draft is wrong. You explain your colors, your logo, where the price goes (bottom, not top). Twenty corrections later you have something usable.
Next week you need another proposal. Claude remembers none of it.
A skill packages all of those preferences into one markdown file that the agent reads before doing the task. Right before recording, Remy typed "create a proposal for Greg Isenberg to sponsor the newsletter" and got a finished, branded proposal in one shot. No re-explaining anything.
His mental model: picture the most capable employee you've ever hired, sitting at a desk. The computer is your MCPs and tools. The instructions file is what he reads before every task. And behind him is a bookshelf of SOPs. Each book spine has a name and a one-line description. When a task needs one, he pulls the book off the shelf and follows it step by step.
That bookshelf is your skills library.
Remy's math: one good skill saves him about 2 hours a week. Stack 20 of those and you've deleted a part-time job.
The problem: we're in the Microsoft Word era of skills
This is the part of the episode that made me stop and take notes.
AI right now is single-player. Your developer has his own Claude setup. Your marketer has hers. When someone builds a brilliant skill, it lives on their machine and dies there.
Remy compares it to documents before Google Docs. You'd write something in Word, email the file around, and end up with ten conflicting copies and no source of truth. That's exactly what happens when you share skills by zipping them into Slack.
You send the file. Your teammate uploads it. Then you improve your copy, and theirs is instantly stale. They fix a bug in theirs, and you never see it.
He tried the obvious workarounds too:
- Google Drive or Dropbox desktop apps, symlinked into Claude's skills folder. Works if everyone on your team understands folder structures. Breaks the second you hire a junior marketer who doesn't.
- Obsidian vaults. Same problem, plus now everyone has to install and learn Obsidian.
If you've read my breakdown of Greg's episode on turning Claude Code into an AI employee, this is the missing piece that episode didn't cover. One employee is easy. A team of humans sharing one brain is the hard part.
The fix: one GitHub repo, installed as a plugin
Remy's answer sounds technical and takes about ten minutes: push every skill into a GitHub repository, then install that repo as a Claude Code plugin.
GitHub, if you've never touched it, is just Google Drive for code. One folder in the cloud, everyone reads from the same copy, and every change is tracked.
A plugin is just a bundle of skills (optionally with MCPs and hooks) that Claude can install from a repo. What turns a folder of markdown files into a plugin is a couple of small JSON files. Remy has no development background and didn't write them himself. He asked Claude to set it up, and he shares the exact prompt in the episode description so you can do the same.
The install is genuinely simple:
→ Type /plugin in Claude Code.
→ Add your GitHub repo URL as a marketplace.
→ Install the plugins you want. Done.
Remy organizes his "team skills" repo by department: brand, content, newsletter, marketing, finance. His copywriter installs the newsletter plugin but not the finance one. The marketplace is the app store, each department bundle is an app.
Two details that matter:
First, everyone must turn on auto-update. That's what makes this multiplayer. When Remy improves the notion-formatting skill, every teammate's copy updates on its own.
Second, this isn't Claude-only. Remy runs the same plugin in Codex, same skills, same repo. One library, any harness.
And if you're on a Claude enterprise account, you can bake the plugin into everyone's accounts as an organization plugin. Your non-technical hires never touch a terminal.
He lost 150 skills before he learned this
The reason Remy is religious about this setup: about two months ago, Claude deleted his local Claude folder.
150 skills. Over 500 hours of work. No backups.
(I felt this one physically. Ask any founder who's lost a database.)
So now everything lives in the cloud, in two repos. The team repo holds company skills. A second personal repo, "Remy Skills," holds the ones that only make sense for him: an inbox triage skill, a morning brief tuned to his apps, plus downloaded skills he's testing before rolling out to the team. It's his sandbox, and it's also how cloud agents he runs on a VPS can access his skills without his laptop being on.
There's a sharper business point hiding in here, and I think it's the most underrated moment of the episode.
Skills are becoming company IP.
Remy's repo lives inside his GitHub organization. When an employee builds a great skill on company time and later leaves, the company keeps the skill. Same way your SOPs don't walk out the door with the ops manager who wrote them.
His bet, and I agree: well-built skill libraries will show up in acquisition due diligence the same way documented processes do today. If you train agents to do real, economically valuable tasks, that training is an asset.
Skill chains, usage stats, and skills that improve themselves
Once the library existed, Remy started building on top of it. Three ideas worth stealing.
Skill chains. His YouTube publishing workflow is one orchestrator skill that calls three smaller ones in sequence: titles, thumbnails, descriptions. His rule for when to split a process into a chain: if you'll ever want to run a sub-step on its own (just thumbnails today, please), make it its own skill. Otherwise keep it as one.
Usage tracking. A hook in the plugin tallies every skill run, per skill and per team member. Half the motivation was a joke (a Spotify Wrapped for skills: "you ran your proposal skill 56 times this year"). The serious half: after 6 months he can see which skills nobody touched and cull them before the stack bloats. He even built a small web app over the repo that visualizes how skills connect, with an assistant that answers "I finished the weekly research, what skill do I run next?"
A self-improvement loop. At the bottom of every skill.md, Remy appends the same block: after each run, review whether any step failed, whether the user corrected anything, and whether a future run needs something you just learned. If yes, propose an update to the skill.
This turned his whole team into skill maintainers without anyone trying. His teammate Harold hit an image-upload error while staging an email in Resend, found the workaround, and updated the email-formatting skill. Now everyone's next run includes Harold's fix automatically.
One person debugs once. The whole company gets smarter.
That compounding effect is the same reason I'm bullish on agent-run business automations in general: you pay the setup cost once, and the payoff repeats every week.
How to start skillsmaxxing this week
Remy's operating philosophy, borrowed from Russ: thin agents, thick skills.
Keep your agent instruction files lean. Put all the context, preferences, and step-by-step detail into fat, specific skills. Then any harness (Claude Code today, whatever wins next year) can execute your playbook.
His trigger for creating one is dead simple: any time he finishes a task with Claude that he knows he'll do again, he types "turn this into a skill." That habit is how he ended up with hundreds.
If I were starting from zero, here's the week-one plan straight from the episode:
→ Pick 3 repeatable tasks you did this week (a proposal, a formatted doc, an outreach email).
→ Do each one with Claude, correct it until it's right, then say "turn this into a skill."
→ Push the skills to a GitHub repo and ask Claude to make it a plugin (Remy's prompt is in the video description).
→ Have your team install it via /plugin with auto-update on.
→ Paste Remy's self-improvement block at the bottom of every skill.
You don't need to start from scratch either. Remy runs Corey Haynes's marketing skills and the Boring Marketer's skill pack as foundations, then layers his own on top.
And if you want to see where this scales to, read my breakdown of how Allie K. Miller runs her company with 34 AI agents. Same thesis, bigger machine. Skills are the SOPs; agents are just the employees who read them.
FAQ
What is skillsmaxxing?
Skillsmaxxing is Greg Isenberg's term for aggressively building and refining agent skills: reusable SOP files that teach AI tools like Claude Code exactly how to do your repeatable tasks. Instead of re-prompting from scratch, you invest in a library of skills that produce your preferred output in one shot, then share and improve them over time.
What's the difference between a skill and a plugin in Claude Code?
A skill is a single markdown file describing how to do one task (write a proposal, format a doc, triage an inbox). A plugin is a distributable bundle of skills, and optionally MCPs and hooks, installed from a GitHub repo through the /plugin command. The plugin is how a team shares one central skills library with version control and auto-updates.
Do Claude Code skills work with Codex?
Yes. Remy runs the exact same GitHub-repo plugin in both Claude Code and Codex, and the skills behave the same way. That's a core part of the "thin agents, thick skills" idea: keep the intelligence in portable skill files so you can switch harnesses without rebuilding anything.
How many skills should a founder build?
Remy's rule: every repeatable, non-physical task in your business can become a skill, and one good skill saves roughly 2 hours a week. Start with 3 tasks you repeat weekly, build those, and add a new skill every time you catch yourself doing something with AI a second time. Track usage and cull skills you haven't run in 6 months.
Steal playbooks like this every week
I break down what actually works for bootstrapped founders (real numbers, real systems, no guru fluff) on the Profitable Founder Podcast. Weekly interviews with SaaS founders doing $100K to $10M a year.
