OpenAI publishes a guide to make Codex perform beyond a prompt

OpenAI and Jason Liu released a guide on using Codex as a persistent workspace that maintains context and memory over time for advanced development projects.

OpenAI publishes a guide, signed by creator Jason Liu, on using Codex for work that extends beyond a single prompt. The central idea: Codex, designed for code, also becomes a persistent workspace, capable of maintaining context and advancing projects over time, for developers and beyond.

The pattern consists of a few building blocks. Work continues as long as Codex has a durable conversation thread, memory kept outside the conversation, tools, a way to resume work at pace, and a place to review the output. The guide details several practices: steering, which involves inserting the next instruction while Codex is working; memory stored in files that can be opened, edited, and compared, a vault distinct from the code repository; remote control to monitor a long task from another device; automations that reactivate the same thread at regular intervals; and verifiable objectives, a good goal providing a success criterion, such as tests to be rerun, rather than a simple plan to execute.

The guide emphasizes human review, with the user retaining approval, tone, timing, and any irreversible action, and points out a trade-off: a long thread carries context and costs more than a new thread. OpenAI sees this as a pattern that extends beyond just code.

The guide is available here.