How would you design production alerts that are useful and not noisy?

In production I review alert performance after incidents: whether it fired early enough, represented real user impact, reached the team that could act, and linked to a useful runbook. Thresholds account for normal traffic cycles, maintenance, and minimum request volume so a tiny sample does not create a page. Alerts connect to dashboards, recent deployments, dependency state, and a named service owner. I track false positives, missed incidents, acknowledgement time, and time to mitigation, then remove or demote alerts that repeatedly produce no action. The goal is a small trusted paging surface, with lower-severity signals available for investigation without waking people unnecessarily.