The Codex feature that works while you sleep

What It Covers

Claire Vaux walks through Codex's /goal feature, which enables AI to run autonomously for hours without human prompting. She covers the six-part goal-writing framework, demonstrates three real use cases — error elimination, inbox cleanup, and task management — and explains when goal-based loops outperform standard turn-based prompting.

Key Questions Answered

• •Goal vs. Prompt Structure: Standard prompting is turn-based — the AI completes one step and waits. The /goal command creates a continuous work-verify-iterate loop where Codex keeps running until it gathers measurable evidence of completion, then reports back. This enables multi-hour autonomous sessions without manual "keep going" prompting between each step.
• •Six-Part Goal Framework: Effective Codex goals require six components: a defined outcome, a verification method, constraints on what cannot regress, boundaries on which tools and files are accessible, an iteration policy for deciding next steps, and a stopping condition that tells the AI when to surface blockers rather than continue attempting fixes independently.
• •Error Elimination Use Case: Point /goal at error logs in tools like Sentry or Vercel, instruct Codex to categorize each error, fix root causes, and replay historical examples to validate fixes. Claire used this approach to reduce a persistent edit-operation error from recurring daily to zero occurrences, with Codex running several hours to produce a systematic rather than patched solution.
• •Non-Technical Productivity Applications: /goal works beyond coding — Claire reduced approximately 3,900 unread Gmail messages to 68 requiring attention in a 3-hour-52-minute session. Codex categorized emails, clicked unsubscribe links, and created labeled folders. A similar approach cleaned hundreds of stale Linear project tasks by applying a consistent rule: cancel any incomplete pre-current-week episode work.
• •When Not to Use /goal: Avoid /goal for single-line edits, vague outcomes like "improve code quality," or tasks without a measurable finish line. The feature is strongest when three conditions exist: a durable objective that stays stable over time, an evidence-based completion condition that can be tested programmatically, and a path requiring multiple investigative iterations to reach.