OpenAI's Codex can now watch you do a task once and turn it into a reusable skill

OpenAI has shipped a feature called Record & Replay for its Codex agent, letting you demonstrate a workflow once on a Mac and have Codex turn that recording into a reusable automation it calls a "skill." The feature landed on June 18, 2026 in version 26.616 of the Codex macOS app, according to coverage from The Decoder and confirmed in OpenAI's own developer documentation.

What it actually does

The idea is to capture tasks that are easier to show than to describe in a prompt. You open the Plugins menu, pick Record a skill, give Codex some context, and grant it permission to observe your screen. Then you just do the task — file an expense, publish a video with its thumbnail and subtitles, book a parking spot — and stop the recording from the menu bar, an overlay, or by saying so out loud.

From there, Codex's model inspects the captured actions and window content and drafts a written skill: a document that spells out when to use the workflow, what inputs it needs, the step-by-step procedure, and how to verify the result. Next time, you trigger the skill and Codex runs the same flow with new inputs, using whatever tools are available in the current environment — Computer Use, browser actions, installed plugins, or a mix.

Because the generated skill is human-readable, you can ask Codex to refine it: clarify a naming convention, pin down a hidden preference, or spell out a decision point it guessed at during recording.

Requirements and limits

Record & Replay is macOS only and depends on Computer Use being enabled — in organizations that use a requirements.toml, the [features].computer_use setting gates access to it. It's available to paid ChatGPT plans (Plus, Pro, Business, Enterprise, and Edu), but OpenAI notes it is initially unavailable in the European Economic Area, the United Kingdom, and Switzerland. The-decoder reports Computer Use itself only became accessible in the EU on June 16, two days before this launch.

Why it matters for people learning to build with AI

The interesting shift here is in how skills get authored. Most agent tooling in 2026 — from Claude Code's SKILL.md files to custom slash commands — expects you to write the instructions by hand. Record & Replay inverts that: you perform the task and the model writes the procedure for you, then hands you an editable artifact to correct.

That lowers the barrier for non-developers to build their own automations, but it also raises familiar questions. A recorded skill is only as reliable as the demonstration it learned from, and a workflow that drifts — a renamed button, a new pop-up — can break a replay in ways a hand-written, well-scoped skill might survive. The honest framing is that recording gets you a first draft of an automation fast; turning that draft into something you'd trust to run unattended is still the part where understanding what the agent is doing pays off.

For anyone learning to build with AI agents, it's a useful reminder that "skills," "slash commands," and "recordings" are converging on the same underlying idea: a reusable, inspectable instruction file that the agent loads only when it's relevant.