Persistent autonomous work-until-done. Register the condition as Claude Code's native /goal <condition> directive. /goal installs a session-scoped Stop hook that BLOCKS the session from stopping until the condition holds true (the evaluator judges surfaced conversation evidence), and auto-clears once met. The condition itself is the directive — keep working toward it, do not pause for instructions.
Periodic progress notifications (~5 min). A heartbeat update on a fixed cadence (default 5 minutes) so the user knows it is still alive even when nothing newsworthy happened.
Ping on each update. Whenever a meaningful state change occurs (a step completes, a test flips, a blocker appears), fire an immediate visible macOS notification — independent of the heartbeat.

This skill is the HOW that wires (2) and (3) onto (1). The bare /goal command gives you the work-until-done loop; it does not, on its own, send periodic or per-update desktop notifications.

<when_to_use> Trigger when the user expresses a durable, condition-gated work intent plus (often) a notification cadence:

"set a goal: " / "goal mode"
"work until " / "keep working until " / "don't stop until "
"notify me every minutes" / "ping me on updates" / "tell me when something changes"
"show me a notification on my computer when …"

Also use when the user wants the agent to drive toward an outcome autonomously (no "what should I do next?" check-ins).

goal-mode