Codex Record & Replay is one of those features that sounds small until you translate it into real work. You can show Codex how you complete a repeated workflow, then Codex turns the demonstration into a reusable skill it can apply next time.
That matters because many business workflows are not hard because the steps are technically complex. They are hard because the preferences are hidden: where the metadata lives, which field names matter, which asset folder to use, how to name things, what should be saved as private, and what counts as "done."
What shipped
OpenAI's Codex documentation says Record & Replay lets you demonstrate a workflow on your Mac and turn it into a reusable skill. It is meant for workflows that are repetitive, preference-heavy, or easier to show than to describe in a prompt.
The official examples include filing an expense, booking a parking space, creating a correctly configured issue, publishing a video, and downloading a recurring report. In the official video, the workflow is YouTube publishing: pull metadata from a spreadsheet, find matching assets, add a thumbnail and captions, save the video as private, and verify the upload.
Important availability caveat: OpenAI says Record & Replay is available on macOS, requires Computer Use to be available and enabled, and initial availability excludes the European Economic Area, the United Kingdom, and Switzerland. For readers in Portugal, that means this is a feature to track now, but not one to assume you can use locally today.
Why it matters
The old automation question was: can I describe this process well enough for an agent to do it?
Record & Replay changes the question to: can I demonstrate the process once, then refine the captured skill so the agent understands the repeatable pattern?
That is a big difference for operators. Many useful workflows are full of little decisions that never make it into documentation:
- Use this spreadsheet, not the old one.
- Find assets by package name, not by upload date.
- Always add English captions before saving.
- Save as private first, never public.
- Verify the result by checking these fields after save.
Those preferences are often why automation fails. Record & Replay gives Codex a way to observe them, then package them into reusable context.
How it works
The workflow is simple enough:
- Open Plugins in the Codex app.
- Use the plus menu and select Record a skill.
- Review the suggested prompt and give Codex context about the goal.
- Approve the recording request when you are ready.
- Complete the workflow on your Mac.
- Stop recording from the menu bar, overlay, or by telling Codex you are done.
- Let Codex inspect the recording and draft the skill.
The generated skill should explain when to use the workflow, what inputs it needs, what steps to follow, and how to verify the result. You can then ask Codex to refine it further.
In practical terms, the skill should end up looking less like a vague transcript and more like an operating procedure:
When to use:
- Use this skill when publishing a prepared video package.
Inputs:
- Video file
- Thumbnail
- Caption file
- Metadata row or package name
Workflow:
- Match the package to the metadata source.
- Fill the required fields.
- Attach thumbnail and captions.
- Save as private.
Verification:
- Confirm title, description, thumbnail, captions, privacy setting, and saved status.That verification block is the part I would obsess over. A workflow skill without a clear "done means..." section is just a softer prompt.
What to record first
Start with workflows that are boring, stable, and repeated. This is not the place to begin with a messy one-off strategy task.
- Publishing: upload a YouTube video, podcast episode, portfolio item, blog post, or newsletter issue.
- Admin: file expenses, download reports, reconcile a weekly spreadsheet, update a CRM field.
- Project work: create a correctly configured Linear issue, format a pull request, open a release checklist.
- Client operations: prepare a recurring client folder, rename assets, add links to a handoff doc.
- Calendar: create invites with your preferred naming, agenda, location, and reminder defaults.
The best first target is a task you already know well enough to notice mistakes. If you cannot judge the result, do not use it as your first Record & Replay workflow.
This is not just a macro
The most interesting detail from the analysis video is that replay does not have to mean "repeat every click exactly." A skill gives Codex reusable context, but Codex can still choose the best available tool in the current environment.
That means a recorded browser task might replay partly through browser actions, partly through connected plugins, partly through scripts, and partly through Computer Use. In some cases, Codex may improve the workflow by using a deterministic script or connector instead of slow UI clicking.
That is where this becomes serious. The recording captures the human process. The skill gives Codex the operating pattern. The replay can use better machinery than the original demonstration.
Limits and caveats
The feature is powerful enough that the guardrails matter.
- Region and platform: macOS at launch, and not initially available in the EEA, UK, or Switzerland.
- Sensitive data: do not record passwords, client secrets, private keys, payment data, or anything you would not want in a skill draft.
- Scope creep: stop recording when the workflow is complete. Do not wander into unrelated cleanup.
- Unstable UI: if the app changes every week, record the durable decision logic, then refine the skill toward robust tool use.
- Team distribution: if you want a stable package for a team, OpenAI points builders toward plugins rather than only a recorded skill.
My rule: Record & Replay is excellent for capturing a personal or team workflow. A plugin is better when the workflow becomes a product, a team standard, or something with install metadata and integrations.
Recording checklist
Before recording, write this down:
- Goal: what should be true when the workflow is finished?
- Inputs: which files, rows, dates, links, or IDs change each time?
- Preferences: what naming conventions, defaults, or hidden rules matter?
- Safety: what should Codex avoid seeing or changing?
- Verification: what fields, files, statuses, or screenshots prove it worked?
- Replay surface: should Codex use Computer Use, browser actions, plugins, scripts, or a mix?
- Human review: where should Codex stop and ask before publishing, sending, deleting, or charging anything?
This is the useful direction for agent work: fewer giant prompts, more reusable skills, clearer preferences, and better proof that the task actually finished.