Claude Code's new /goal command enables a true agentic loop where the model iterates autonomously until a specified goal is met. A separate evaluator model judges each turn's output, preventing Claude from grading its own work. The command works best with measurable, verifiable goals (e.g., all tests passing, zero lint errors), a CLAUDE.md file describing project context, and optionally auto mode for fully unattended runs. Every action is logged, and Ctrl+C stops execution at any time. The author argues this is the first Claude Code feature that genuinely earns the 'agentic' label.
Table of contents
What /goal actually does (and doesn't do)The setup that makes this work bestLet Claude cook2.3K Impressions