Codex's /goal command turns the AI from a turn-by-turn assistant into an autonomous agent that runs multi-step tasks for hours without supervision. In a 30-minute episode, the host demonstrates three real runs: a five-hour-and-45-minute session eliminating hundreds of Sentry and Vercel API errors in a live codebase, an inbox reduction from 3,900 emails to 68 completed in under four hours, and a bulk cleanup of hundreds of Linear project tasks.
The episode is worth watching for the structure it provides around Goal-writing, not just the demos. At the 7:34 mark, six specific components of an effective /goal prompt are outlined, covering measurable outcomes, verification methods, and hard constraints. The distinction between outputs and outcomes, covered at 6:06, is what separates Goals that run to completion from Goals that stall or drift.
The episode also covers when not to use /goal, which is the section most people will skip and shouldn't. The framing throughout is a direct shift in mental model: from babysitting a model prompt-by-prompt to managing an agent by objective. If you are still treating Codex like a chatbot, this episode explains what you are leaving on the table.
[READ ORIGINAL →]