Hello and welcome back to my newsletter!

I often get asked: "Which AI do you use? ChatGPT or Claude?"

The slightly more technically advanced will ask: "What is your harness? Codex? Claude Code? Pi?"

The questions all expose one huge misconception when working with AI Agents:

The agent is not the unit of value. The agent is the interface. It is the thing you talk to. The thing that decides which capability to invoke.

The capability is the unit of value. The biggest value lives in the skills you teach your agents.

A skill is a small, focused capability you teach an agent. Think of it as a recipe. It tells the agent how to do one specific thing: publish a blog post, send an invoice, mirror a website, look up a customer in the CRM. Each one bundles a description, the steps to follow, and sometimes a small script that does the actual work. The agent reads the description, decides whether the skill fits, and follows the recipe.

If skills are where the value lives, then managing skills well becomes a craft of its own.

You have to organize them. You have to give each agent access to the right ones. Secrets need to stay where they belong. And you have to stop the same skill from drifting into three half-broken copies.

This is the part most people skip. They build a skill, drop it wherever the agent expects it, and move on. Then the mess starts.

Wrong agent. You taught the skill to one tool, not the others. The workflow only runs in one place. The other agents are useless for that job.

Conflicting versions. Three copies of the same skill, all slightly different. The fix you made last week only lives in one of them.

Secret drift. You rotate an API key in one folder. The other copies still hold the old key, fail silently, and you spend an hour debugging the wrong thing.

Diverging state. A skill that issues invoice numbers from a local counter hands the same number to two different clients, because the counter exists in three places at once.

All four have the same root cause. The skill is being treated like a setting on the agent, instead of an asset you own.

How to fix it?

You stop putting skills inside the agents.

You keep your skills in one place. One canonical repository on disk. Each agent points at that repository through symlinks. The agent thinks it has its own skill folder. In reality, every entry is a link back to the library.

Update a skill once. Every agent picks up the change. Rotate a secret once. Every agent uses the new value. Add a new agent next month. Point it at the same library. It inherits everything you have built so far.

Three copies become one. Three things to maintain become one. Three half-broken versions become one good version.

The agents stop feeling like islands. They start feeling like different ways of asking the same library to do work.

That is what changes when you treat the skill as the unit and the agent as the interface.

How to get started

This week's subscriber post for my Explorers has the full setup. Every command. The installer script. The promotion script. The secret-loading pattern. The troubleshooting guide. A worked example of taking a real skill from prototype to canonical.

If you want to build the same thing on your machine, that is where the complete walkthrough lives.

→ Read the full guide (Explorers only)

Have a great day!

Matthias