The /goal slash command in 📝Hermes Agent locks the agent on a user-defined completion condition across turns until a separate judge confirms it has been met, shipped in v0.13.0 as Hermes' canonical implementation of the 📝Ralph Loop pattern.
The command accepts a natural-language target — "send ten outreach emails," "produce a five-slide HTML presentation on the top Product Hunt launches" — and the runtime iterates the worker against a separate judge call until the criterion is satisfied or the user cancels. Like every Hermes slash command, /goal is invocable identically from the terminal UI and from any platform connected via the Hermes Gateway, so a goal initiated in Slack or Telegram advances against the same agent state as one started locally.
Effective /goal use depends on the three gating heuristics that govern the Ralph Loop generally: completion conditions must be measurable so the judge can evaluate them, scoped to roughly twenty turns so the loop converges, and self-served so the worker does not stall waiting on human input. Hermes' implementation extends the basic pattern by stitching /goal into the rest of the runtime. The judge can read from FTS5 session search and persistent memory; the worker can spawn Kanban subagents for parallelizable subtasks; and the trajectory left behind is candidate material for the Autonomous Curator to distill into a reusable agentskills.io skill document. The cumulative effect is that a single goal often leaves behind a reusable skill the next similar goal loads rather than re-derives.