A developer using a coding agent, mid-session, faces current context plus a set of available skills and must decide which to invoke next without picking wrong.
Today people name the skill in the prompt by hand or let the model decide, with no separate ranking or abstention mechanism.
Once skills multiply, agents pick wrong or pick everything, forcing developers to correct repeatedly and burn context turns; public material only shows the ranking and abstention mechanism exists, not user complaints or adoption data.