OpenAI has published a practical guide on how to use its coding assistant Codex not just for one-off prompts, but as a "workspace" that sustains work over long periods[1]. Written by Jason Liu of the Codex team, it lays out concrete approaches for preserving context across projects that run for days, managing complex workflows, and keeping progress moving. The guide reframes the AI from a mere chat partner into a place where work actually happens.
The Goal: Work That Extends Beyond a Single Prompt
The guide starts from the observation that organizations are increasingly using AI for work that does not finish in a single exchange[1]. Codex spread first as a tool for writing code, but this guide pushes its use toward "long-running work."
Specifically, it rests on three pillars: breaking ambitious goals into verifiable steps, maintaining continuity across multiple workstreams, and judging when execution should be delegated to Codex versus when human oversight matters most[1]. The more ambitious the goal, the more it depends on a design that breaks the work into stages so the direction does not drift along the way.
"Ambition Without Verification Is Just a Wish"
What Jason Liu emphasizes is that having a long conversation to build up a plan and then saying "implement this" is a different thing from drawing out good execution. The quality of execution is only as good as the goal and the verification you give it. He puts it bluntly: ambition without verification is just a wish[2].
For that reason, he recommends giving work an "operating loop." Concretely, that means a durable thread, shared memory, tools that can actually act on the computer, ways to steer and resume the work, and a surface to review the artifact itself[2]. If you place an idea in a thread, give it memory, and attach verification criteria, you can keep the work progressing without watching it constantly, just checking the diff occasionally and correcting the direction.
The Side Panel That Turns Codex Into a Workspace
A key element in the guide is Codex's side panel. According to Liu, the side panel is what turns Codex from merely a chat app into the place where work happens, and it serves three jobs: inspecting the artifacts that are produced, operating web surfaces, and reviewing changes[2].
Because you can layer instructions while looking at the artifact and, when needed, step in to operate or fix it on the spot, the back-and-forth between dialogue and real work is completed in one place. For work that runs long, how easily you can spin this "review and fix" cycle determines efficiency.
Beyond Coding, Into Knowledge Work
Liu recalls that around late 2025 he began pushing coding agents into knowledge work as well[2]. For example, he made presentation slides with a slide tool and, combined with voice input, used an agent like a note-taker, exploring the range of artifacts a coding agent could produce.
OpenAI's guide likewise positions Codex as a "persistent workspace that preserves context, manages complex workflows, and helps sustain progress across long-running projects"[1]. At the same time, it explicitly notes the importance of judging how much to delegate to Codex and where human oversight is most valuable, assuming a style of use in which humans keep the initiative on judgment rather than handing everything off.
Summary
The guide OpenAI has published lifts users' thinking from using Codex for one-off prompts to wielding it as a "workspace" for running long-horizon work. By breaking big goals into verifiable steps and combining a durable thread and memory with a side panel for operating and reviewing, it shows an approach in which work keeps advancing while humans simply check in from time to time. As use spreads beyond coding into knowledge work in general, how you design AI as a "place where work happens" looks set to be the dividing line for getting the most out of it.
