Claude Code's /goal Command

Claude Code's /goal command lets you define a completion condition once, then Claude keeps working autonomously across multiple turns until a separate evaluator model (Claude Haiku by default) confirms the condition is met by reading the conversation transcript. This removes the need to manually prompt 'keep going' during long tasks like migrations or test suite fixes. The evaluator only judges what Claude has surfaced in the transcript, so conditions must describe observable, verifiable end states. The post covers how the evaluator loop works, how to write effective conditions, the three execution modes (interactive, non-interactive with -p flag, and remote control), how /goal compares to /loop, Stop Hooks, and Auto Mode, and its limitations including no built-in token budget and inability to handle vague or compound objectives. Best practices include using CLAUDE.md for context, PostToolUse hooks for auto-validation, and combining /goal with Auto mode for unattended runs.