OpenAI Codex adds Record & Replay for reusable workflow skills

• OpenAI added Record & Replay to Codex, and OpenAIDevs' rollout note says it turns a demonstrated workflow into an inspectable, editable skill.
• The core interaction is demonstration-first: gdb's post framed it as teaching Codex by example, while OpenAIDevs' demo thread used expense reports and time-off requests as the canonical repetitive tasks.
• Availability is narrow on day one. aibuilderclub_'s Mac app post said it is live in the Mac app, and testingcatalog's rollout note said the feature is not yet available in the EEA, UK, and Switzerland.
• OpenAI had already been testing adjacent skill-teaching mechanics. mattpocockuk's /teach post said /teach creates reusable components, and mattpocockuk's Rubik's cube example showed those taught skills persisting across later use.

You can check OpenAI's rollout post, skim the /teach quickstart, and watch the Rubik's cube proof clip from the earlier skill-teaching flow. OpenAIDevs also spelled out the product shape in one sentence: you choose when recording starts and stops, then Codex turns the session into a reusable skill.

The new feature sits on top of Codex's computer-use stack. According to OpenAIDevs' demo thread, you perform a recurring task once, explicitly start and stop the recording, and save the result as a skill that can be inspected and edited before reuse.

The strongest detail in the launch copy is the handoff from recording to execution. testingcatalog's post described the output as an executable skill, not a one-off macro, which matches OpenAIDevs' wording about reusable skills rather than a transient recording.

OpenAI's examples stayed close to routine desktop work: expense filing and time-off requests in OpenAIDevs' thread, then broader repetitive workflows in aibuilderclub_'s Mac app summary.

Under the hood, the saved skill can be replayed with new inputs through multiple surfaces, according to aibuilderclub_'s post:

• Computer Use
• Browser actions
• Plugins

That multi-surface replay claim is the interesting part. It suggests the recorded workflow is stored as a higher-level skill object, not just a raw clickstream, because the post explicitly says the same skill can later run with new inputs.

Day-one access is constrained in two ways. aibuilderclub_ said the feature is Mac-only for now, while OpenAIDevs said Record & Replay is rolling out only to select markets before a broader release.

Community posts filled in the geography. testingcatalog's note said the feature is not yet available in the EEA, UK, and Switzerland, and koltregaskes' follow-up highlighted the same regional gap alongside the lack of Windows support.

That timing stands out because OpenAI had just expanded other Codex features into those regions. testingcatalog's earlier regional rollout post, dkundel's recap, and WesRoth's summary all described computer use, Chrome extension support, memory, and Chronicle reaching the EEA, UK, and Switzerland days earlier.

Record & Replay looks like the productized successor to /teach. Earlier in the week, mattpocockuk's post said /teach creates reusable components that improve as you learn more, with a side claim that the approach is more token-efficient.

The earlier /teach flow also answered a skepticism question that matters here: whether taught behavior persists. mattpocockuk's Rubik's cube example argued it does, then linked the saved skill and showed the result in a follow-up video. Record & Replay shifts that idea from command-driven teaching to a visible record-once workflow UI.