Most people keep their AI skills locked away on their personal laptops. Not Remy. He packaged his entire company's AI skills into a single GitHub repository that every teammate installs once, then gets automatic updates forever and can roll back to any previous version at any time.

Remy is the host of AI with Remy, and he recently shared this workflow on The Startup Ideas Podcast. His entire company runs off this one GitHub repo — all team skills are organized by department, and distributed as a plugin to every person's Claude and Codex instances.

Let's start with two quick definitions. A *skill* is an SOP written for AI, stored as a markdown file that outlines exactly how to complete a specific task. Examples include "write a sponsorship proposal" or "format a Notion document". A *plugin* is the packaged format for these skills that Claude Code and Codex can read natively.

The core problem Remy solved: Most people use AI like a single-player offline game. Every time you write a proposal, you have to re-explain where the logo goes, what colors to use, where the pricing belongs. Remy saved this entire process as a skill, so next time he just says "write a sponsorship proposal for Greg Eisenberg", and Claude automatically pulls in the right skill to output a document that matches the company's brand formatting.

Remy built a handful of skills in just one week: Notion formatting, brand voice guidelines, email formatting for Resend. Each skill saves 2 hours a week, and the compound gains add up quickly.

But he ran into a practical problem for team collaboration: Everyone has their own Claude or Codex setup. If someone builds a great new skill, no one else can access it easily. Share it as a zip file? You end up with version chaos. Store it on Google Drive? Claude only reads from its local .skills folder, which requires symbolic links that non-technical team members can't figure out.

Remy's solution: Store all skills in a GitHub repository and package it as a plugin.

GitHub does for skills what Google Drive does for documents. Remy created a repo called Team Skills, organized by department: Brand, Content, Communications, Marketing, Finance, General. This becomes the single source of truth for all of the team's AI skills.

Under the hood, the plugin is just a collection of JSON files that Claude and Codex can recognize. Remy doesn't even know JSON — he had Claude write all of it for him. He has no development background, and it still worked perfectly for him.

Installation is simple: In Claude Code, type /plugin, go to marketplaces, add a new marketplace, and paste the repo URL. Since skills are split by department, writers don't have to install finance-related skills. Turn on automatic updates, and if anyone improves a skill, the whole team gets the updated version automatically the next time they run it.

Codex also supports this same plugin setup. Remy runs the exact same set of skills in both Claude and Codex. Enterprise Claude even offers organization-level plugins: the CTO sets it up once, and it works for everyone, so non-technical employees never have to touch the terminal.

GitHub brings free extra benefits out of the box: version control, so if you mess up a skill you can roll back; company ownership, since the repo belongs to the organization, skills stay even when an employee leaves; and continuity, so institutional knowledge doesn't walk out the door with people.

Remy also has a separate private repo for personal skills: inbox sorting, morning briefings. Why store them separately? It gives him cloud access, so agents running on his VPS can also call the skills. Plus, version control. He learned this the hard way: two months ago, Claude accidentally deleted his .claude folder, which held 150 skills and over 500 hours of work, all with no backup. The repo acts as both his working system and a safety net.

When it comes to design, Remy's core principle is this: Any online process can be automated, as long as you write a good skill or chain of skills. The only exception is physical tasks, like washing dishes. Every time he finishes a repetitive task, he says to himself: "I should turn this into a skill."

When should you split a skill into a chain? When the sub-processes need to be used individually. For example, his YouTube publishing process is an orchestration skill that calls three sub-skills in order: title, thumbnail, description. Sometimes he only needs to generate a thumbnail, so splitting into a chain makes sense.

His other principle: Thin agent, thick skills. Keep your core agent instructions lean, and make your skill files thick and packed with specific details. This lets any tool produce the output you want. (Remy credits this idea to another creator, whose name was not clearly captured in the podcast transcription.)

Skills can even improve themselves. Remy adds a self-audit block at the bottom of every SKILL.md: when the skill finishes running, it checks itself — did any step fail? Did the user correct anything? Did we uncover any information that will be useful for future runs? It then proposes edits, which only go live after Remy approves them.

The system is already being used by the whole team. A coworker named Harold ran into an error when uploading images to Resend, figured out a workaround, updated the existing skill, and pushed the change. Remy got the improvement automatically, and now the whole team uses the better version.

Remy also built a simple web app on top of the repo, similar to what Supabase is for backends. The interface displays a skill dependency graph, an editable skill library, and a Q&A box — you can ask things like "what skills do we have for making decisions?" or "I finished my weekly communications research, what skill should I run next?" It answers based on your private skill library and tells you what to run next. He also added a usage tracker that logs who ran each skill and how many times. Down the line, this could turn into a company version of Spotify Wrapped, flag skills that haven't been used in six months, or surface skills you should be using but haven't installed yet.

Action Checklist:

1. Pick three repetitive tasks you completed this week, turn each into a skill

2. Create a Team Skills repo in your GitHub organization, organize into directories by department

3. Ask Claude to generate the required JSON files to turn your repo into a plugin

4. Add the marketplace in Claude Code and install the plugin

5. Turn on automatic updates for everyone

6. Create a separate private repo to store your personal skills and downloaded skills

7. Add the self-improvement loop block at the bottom of every SKILL.md

Remy's core takeaway: Skills are SOPs, and SOPs belong to the company. Most people treat AI outputs like a lottery ticket: they start a new chat, re-explain their preferences, and accept a random result. That's where low-quality AI output (what people call "slop") comes from.

The chain of logic works like this: Repetitive task → Skill → Repo → Plugin → Whole team → Improvement loop. What you're really doing is storing your collective judgment. Every skill is a record of "what good looks like" for that task — write it once, run it forever.

Remy says: "I don't think people are building nearly enough skills, and they're not building them nearly well enough."

Listen to the full episode on the podcast: Apple / Spotify / YouTube.