If you have ever tried to make AI do your actual job, you have probably run into the same wall: you know the task cold, but writing it out as a prompt feels like translating it into a language you do not speak. You can do the work in your sleep. Explaining it to a machine is the hard part.
On June 18, 2026, OpenAI shipped a feature built for exactly that wall. It is called Record & Replay, and the idea is simple enough to explain in one sentence. You do a task on your Mac one time while the tool watches, and it turns what you did into something it can repeat for you later.
For a business owner who wears every hat and has no IT department, this is worth understanding, even if you never touch it today. It shows where the work is heading, and it quietly proves a point about how to get real value from AI.
What Record & Replay actually is
Let's get the details right, because the early chatter has been muddy.
Record & Replay is a feature of OpenAI's Codex app on macOS. It is not a button inside the ChatGPT chat window. You reach it on a paid ChatGPT plan through the Codex app, not the chat box you already use, with Computer Use turned on. It runs on Mac only for now, and at launch it is not available in the European Economic Area, the United Kingdom, or Switzerland.
At launch, Record & Replay runs on macOS through the Codex app, and is not available in the European Economic Area, the United Kingdom, or Switzerland.
Source: OpenAI Codex documentation, 2026OpenAI describes the feature plainly: "Record & Replay lets you demonstrate a workflow on your Mac and turn it into a reusable skill." It is built for tasks that are repetitive, that depend on your personal preferences, or that are easier to show than to describe in words.
Here is how it works in practice. You start a recording, then you do the task yourself. During that recording, in OpenAI's words, "Codex observes the actions and window content needed to learn the workflow." When you are done, it "drafts a skill," which Codex stores as a reusable set of instructions you can run again without recording it a second time.
The examples OpenAI gives are the kind of small, draining tasks every owner recognizes: file an expense, book a parking space, create a correctly configured issue, publish a video, or download a recurring report. Once the skill exists, you can replay it with browser actions, connected plugins, or a combination, and you do not have to demonstrate it again.
One technical note for the curious: the feature relies on a capability called Computer Use, which has to be available and turned on by you or whoever administers your account. The feature shipped in Codex app version 26.616. One caution worth flagging: OpenAI tells users to keep recordings focused on the workflow and to avoid entering secrets or sensitive data while recording.
Why this is the most literal "show, don't tell" yet
Most AI tools ask you to describe what you want. That is the part that stops people who feel behind. You sit there trying to phrase the perfect instruction, and the friction wins.
Record & Replay flips that. You stop describing and you start doing. The tool learns by watching you complete the work you already know how to do. That is the lowest technical bar AI has offered for building your own automation, and it is a real step toward the kind of self-sufficiency the 7 Levels of AI Proficiency is built to grow.
In the language of The 7 Levels of AI Proficiency, a feature like this can pull someone forward fast. A person who has only ever typed questions into a chat box (early on the climb) can suddenly produce a working, repeatable process by demonstrating it once. The instrument lowers the cost of entry. What it cannot do for you is the part that comes next, and that is where the lesson lives.
The catch nobody put in the headline
Here is the honest part, and it is the most useful thing in this article.
Record & Replay records what you show it. If you show it a clean, well-ordered process, it learns a clean process. If you show it a messy one, with the extra clicks, the workaround you have used for two years, and the step you only half remember, it learns the mess too. A demonstrated workflow becomes an automated workflow, warts and all.
OpenAI says as much in its own guidance. The feature "works best when the steps are stable and the success criteria are clear." The company tells users to record a task they already know how to complete, and to refine the result afterward to capture the hidden preferences a recording might miss, things like naming conventions, default field values, or the points where you make a judgment call.
Read that again, because it is OpenAI telling you to design the process before you automate it. That is the whole argument in two lines. The tool can capture your steps, but it only pays off when those steps are worth capturing.
Design the workflow first, then build the agent to run it.
This is the part LaunchReady has been saying since the start: design the workflow first, then build the agent to run it. An agent inherits your process. If the process is good, the agent multiplies a good thing. If the process is sloppy, the agent multiplies the sloppiness, faster and more often than you ever would by hand.
How to think about it this week
You do not need a Mac or this specific feature to take the lesson. The reps below cost nothing and pay off no matter which tools you end up using.
Pick one task, and pick the right kind
Look for a task that is repetitive, stable, and yours. The recurring weekly report is the classic example, and it happens to be one of OpenAI's own. A task that changes shape every time it runs is a poor candidate for automation, whether you record it or describe it. A task you do the same way every week is a strong one.
Write the success criteria in one sentence first
Before you automate anything, finish this sentence in plain words: this task is done correctly when a specific, checkable thing is true. If you cannot finish it cleanly, the task is not ready to hand off, to a tool or to a person. That one sentence is the difference between automating a result and automating a habit.
Map the steps before you record them
Walk through the task once on paper. Notice the steps you skip, the shortcut you take, the moment you stop and decide something. Those decision points are usually the parts a recording struggles to capture, and they are exactly the parts worth getting right. Clean up the process on paper, then automate the clean version.
The owners who get the most out of tools like this are not the ones with the most technical skill. They are the ones who understood their own work well enough to hand a good version of it to the machine.
A next step
You do not have to adopt a new tool to act on this. Open your week and find the one task you do the same way every time. Write a single sentence describing when it is finished correctly, then map its steps and clean them up. That small piece of work is the thing every automation, recorded or written, is going to inherit. Get it right once, and every tool you try afterward starts from a better place.
Related reading: Level 5: The Captain (Design Thinker).
Sources
- Record & Replay, OpenAI Codex documentation
- Skills, OpenAI Codex documentation
- OpenAI Codex changelog (v26.616)
Frequently Asked Questions
Is Record & Replay part of ChatGPT?
It is a feature of OpenAI's Codex app on macOS, reached through that app rather than the ChatGPT chat window. Paid ChatGPT plans reach it through the Codex app, with Computer Use enabled, not the chat interface they already use.
Does it work on Windows?
At launch, no. OpenAI documents it as a macOS feature. It is also not available yet in the European Economic Area, the United Kingdom, or Switzerland.
What is the biggest mistake to avoid?
Recording a messy process. The tool learns whatever you show it. If the underlying workflow has problems, the automation copies them. Clean up the steps and write down what done right means before you record.
Find your AI Proficiency level
The free 7 Levels assessment places you across seven stages of AI capability. Under ten minutes. Research-backed scoring.