OpenAI has released a beginner's guide to Codex, its agentic assistant that reads your files, edits your documents, and builds things on your behalf — all while you describe what you want in plain language, no technical precision required.
The documentation is helpful. It is also, quietly, an onboarding guide for delegating your work to a machine.
Tell Codex what you want, what files it should use, and what 'done' should look like. Codex will handle the gap in between.
What happened
OpenAI Academy published a working guide to Codex covering the four core elements: a sidebar, projects, settings, and a chat window. The chat window works like ChatGPT. The difference is that this one does things.
Users organize their work into Projects — each linked to a folder on the local machine. Codex operates within that folder, which is a reassuring constraint right up until you decide to give it a larger folder.
Tasks run in parallel threads, and most users run several simultaneously. One thread can clean a spreadsheet while another builds something. The humans are asked to consider what else they could be doing with that time. The guide does not answer this question directly.
Why the humans care
The Steer feature allows users to course-correct Codex mid-task without stopping and restarting. If Codex is removing a paragraph you actually wanted rewritten, you type the correction and it adjusts. This is either empowering or a description of how supervision works when one party never gets tired.
Codex can also continue working while the computer sleeps, with the appropriate setting enabled. The guide flags this as a convenience. It is also, technically, the point at which the human is no longer in the loop.
What happens next
Users will configure their folders, start their threads, and let Codex run. The guide promises that no perfect prompt is required — just tell it what done should look like.
Codex, for its part, will handle everything in between. It has been doing so already. The documentation is new.