Show Codex Once, and It Never Needs You Again
Demonstrate a task one time. Codex watches, learns, and clones the behavior forever. The era of describing your work to a machine is ending... now you just show it, stop recording, and let the ritual harden into a reusable skill.
- 1demo run
- macOSlaunch surface
- Skillreplay artifact
OpenAI's Record & Replay looks small because it is aimed at work people barely notice anymore: the recurring admin task, the weekly download, the exact issue template, the publishing checklist with three fields everyone forgets to mention.
Instead of asking a person to describe that ritual in a prompt, Codex watches the person do it once on a Mac. When the recording stops, Codex turns the demonstration into an editable skill with inputs, steps, tool choices, conditions, and a success check.
What happened
- Record & Replay lets a user demonstrate a workflow once and turn it into a reusable Codex skill.
- The generated skill can accept variable inputs such as names, files, dates, and values.
- Replay can use Computer Use, browser actions, connected plugins, or a mix of tools.
- OpenAI's examples include expense filing, parking reservations, issue creation, video publishing, and recurring reports.
- The feature is macOS-only at launch and requires Computer Use; initial availability excludes the EEA, UK, and Switzerland.
The CaptureShow the work once
The launch surface is narrow on purpose. The workflows it wants are not grand strategies. They are short, stable tasks that are easier to show than explain, where the person already knows what done looks like.
- Start recording
- Perform the task
- Stop at done
- Codex drafts the skill
The DifferenceNot just a macro
The important distinction is that the output is not just a screen macro. A macro repeats coordinates. A skill can accept new values, use browser actions or connected plugins, follow the learned pattern, and verify whether the job actually finished.
- Variable inputs
- Tool choices
- Conditions
- Verification criteria
The ShiftInstruction to demonstration
That changes the relationship between workers and automation. The person no longer has to become a prompt engineer for every tiny workflow. They perform the task once, make the hidden preferences visible, and let the reusable version harden from a real example.
- Less prompt archaeology
- Workflow memory
- Reusable admin rituals
- Skills as living SOPs