The one thing nobody warns you about when you train an AI assistant: it doesn't share notes with itself.
This past week I trained Claude to run a chunk of our CRM workflow — timestamp reconciliation, prorating jobs, invoicing, QBO entry, the works. Each task lives as its own "skill." On one machine, it's been a game changer. Then I tried to use it on a second workstation. And learned the hard way: skills are stored locally. What you teach Claude on one computer doesn't travel to another. So all that careful, step-by-step training I did? It lives on that one machine and nowhere else. My first fix attempt: point both machines at a shared Google file — a single mirrored skill set that every instance could read from and update. In theory, teach it once, use it everywhere. In practice, it's been a headache. Syncing, versions, which Claude updated what and when... it's a real problem. Here's the answer I'm going to try this week: instead of sharing a live folder, package the skills into a plugin and host it in a shared repo, then install that plugin on each workstation. The idea is that every machine pulls from one source of truth — and when I improve a skill, I push the update once and the other stations get it. Version-controlled instead of file-synced. I haven't done it yet, but it looks like the real fix, and I'll report back on how it goes. It's a strange thing to wrap your head around. You're not training one assistant — you're training one per machine, unless you solve the sharing piece. For a small business running multiple stations, that's the difference between "we have an AI assistant" and "this computer has an AI assistant." Honestly if you were to ask my wife she would tell you I might be going crazy arguing and reprimanding the robot that lives in my computer. Still worth it. But if you're going down this road, know the wall is there before you hit it. And if anyone's already cracked clean skill-sharing across workstations, I'm all ears.