Someone on your team finds a free AI add-on on LinkedIn that promises to cut hours off a project. They want to try it on a real client deliverable this week. That's the exaclty moment you need to pause on.

Claude calls these add-ons Skills. ChatGPT has something similar with Custom GPTs and plugins. The name changes, the risk underneath doesn't: you're letting someone else's code or instructions run inside a tool that touches your work and often your client's data.

Why is that risky?

→ Some add-ons are just written instructions. Worst case, the AI gets confused. Nothing runs by itself.
→ Some come with real code attached. That code runs on your computer, and almost nobody stops to read it first.
→ Some connect to other apps in the background, sometimes changing how the tool behaves by default, without you seeing any of it.

Your team member isn't being careless. They're moving fast because that's what they're being asked to do. But "it looked impressive in the demo" and "it's safe to run against client material" are two different things, and nobody's checking for the second one in that moment.

I use skills constantly in my own work, the difference is I read what's inside before I trust one. A skill came across my feed promising to turn a brief into a stop-motion ad automatically, chaining three tools through a connector, "sending it to everyone who asks." I went to check it before trusting the demo and couldn't even get past the CAPTCHA to see what was inside. That alone was enough to leave it alone. 😑

If your team runs something unverified against a client's brand material you're the one explaining that to the client afterward, not the person who shared the free link.

This is exactly the check I run before recommending any AI tool to a company: what it touches, where the data goes, what's running behind the interface. Not an extra step. The job. Of course I coordinate with legal and IT Teams, that need to be involved in these conversations.

‼️ If this is a conversation you haven't had with your team yet, it's worth having before the next "free skill" shows up in someone's feed.

PS: still deep in Claude Code myself this week, turns out reading the security model is half the course.