Codex Turns Recorded Workflows Into Reusable Editable Skills
OpenAI presents Record & Replay in Codex as a way to turn a demonstrated recurring workflow into an inspectable, editable skill. In the source example, a user records a YouTube upload process once, and Codex converts the observed steps, defaults and file conventions into a reusable `SKILL.md`. The argument is that repeat work can move from long prompts and remembered preferences to short invocations, with Codex applying the learned workflow to the next relevant task.
Codex turns a demonstrated workflow into an editable skill
Record & Replay lets a user teach Codex a recurring process by demonstration rather than by spelling out every instruction in a prompt. The user controls when recording starts and stops, performs the task once, and Codex turns what it observes into a reusable skill.
The example is a repeated YouTube publishing workflow. The manual version requires pulling metadata from a publishing spreadsheet, finding the matching assets, and working through the same fields and settings in YouTube Studio each time. Instead of describing that process from scratch, the user starts a recording in Codex with the instruction: “Watch me upload this YouTube video so you can handle these uploads for me in the future.”
After the demonstrated workflow, Codex reports that it “stopped the recording, inspected session.json and events.jsonl, and created a reusable skill at @SKILL.md.”
The on-screen summary names the files Codex inspected — session.json and events.jsonl — and the skill file it created, SKILL.md. It also reports that two files were edited. The source describes the resulting skill as inspectable and editable, though the contents of SKILL.md are not shown.
The recording captures defaults that would otherwise live in the user’s head
The demonstrated task is concrete: upload a YouTube video using information from a publishing calendar and files from a prepared package. On screen, YouTube Studio is open beside a Google Sheets publishing calendar. The visible row includes the title “Introducing Sites in Codex” and a description for that video. As the user performs the workflow, they pull in the title and description, add the thumbnail and English captions, and save the video as private.
Those choices become reusable defaults. The visible Codex summary says the skill captures the YouTube Studio upload flow: prepared folder inputs, checking the publishing sheet first, selecting the .mp4, adding matching .srt subtitles, using “not made for kids” by default, proceeding through checks, saving as private unless explicitly told otherwise, and reporting the final YouTube link or warnings.
The important detail is that the demonstration captures both the visible mechanics and the surrounding conventions: where the metadata lives, how assets are paired, which defaults to apply, and what verification means. Those preferences are not restated as a long prompt. They are inferred from the recorded workflow and written into the skill.
You don't have to explain every step or preference in a prompt. Just show Codex how you do it.
The source describes the skill as editable, and the screen shows that SKILL.md was created after Codex reviewed the recording. It does not display the skill file’s contents.
Replay changes the request from instructions to invocation
After the skill is created, the user opens a fresh thread, attaches the next video package, and asks Codex to apply the learned upload workflow. The prompt shown on screen is short: “Upload this youtube video using @youtube-upload.” The attached files include an MP4 and an SRT subtitle file.
Codex then applies the recorded procedure to the new upload. It matches the package to the right row in the spreadsheet, fills in the metadata, adds the thumbnail and English captions, uploads the video as private, and verifies that everything was saved correctly. The visible YouTube Studio dialog shows the new upload being filled in with the title “Debug web apps with browser use in Codex” and a description about using Chrome DevTools Protocol in Browser Use to inspect console logs, runtime errors, local storage, applied styling, network traffic, and performance profiles while working on web applications.
The burden on the user changes. The user no longer restates the spreadsheet location, the asset-matching convention, the caption rule, the private-visibility default, or the verification step. The request names the skill and supplies the next package; Codex applies the process it learned from the earlier demonstration.
OpenAI gives examples beyond video publishing: filing an expense report, submitting a time-off request, formatting and sharing a pull request, or setting up calendar invites according to a user’s preferences. In each case, the pattern is the same: demonstrate the workflow once, let Codex convert the process into a skill, and invoke that skill later with the relevant context.
When reused, the skill can complete the task through computer use, browser use, connected plugins, or a combination of them. Record & Replay is not described as replacing those capabilities. It packages a demonstrated workflow so Codex can later execute it using the tools available in the environment.