Write skill descriptions as model-facing triggers, not user summaries — load them with concrete action verbs

Anthropic/Claude·medium signal

Anthropic's skills guide stresses that a skill's description is not a user-facing summary; it is the trigger text the model scans to decide whether to load the skill at all. Pack it with concrete action verbs that match how the task actually surfaces ('babysit', 'debug', 'reconcile'), not polished marketing prose. Mismatched or generic descriptions cause the right skill to never fire, which is the silent failure mode of large skill libraries.