Filing a bug issue is easy. Writing down every convention around it is usually the tedious part.
A team may have rules for title prefixes, project assignment, labels, priority and description structure. An experienced operator applies them without stopping to explain each choice. A prompt has to make those choices explicit, which is where small but important rules tend to disappear.
Record & Replay lets you teach that kind of SOP by doing it once. Perform the workflow on a Mac, let Codex observe the actions and relevant window content, then stop the recording. Codex turns the demonstration into an inspectable, editable Skill that can be used again.[1][2][3][7]
The first run is still manual. On the next similar task, reuse the Skill with new inputs instead of explaining the whole SOP again.
What the recording leaves behind
Record & Replay runs on macOS and depends on Computer Use. The user controls when the recording starts and stops; while it is active, Codex observes the parts of the desktop workflow it needs in order to draft the Skill. The initial rollout also had regional limits.[1][2][3][4][7]
Read the draft and check four things:[7]
- When should this workflow run?
- Which inputs can change between runs?
- What steps belong to the procedure?
- How should the result be checked?
A later run can use a different issue, file, date range or other input. Codex then works with the tools available in that environment, including Computer Use, browser actions and installed plugins where available.[3][4][7]
The recording itself is also input data. OpenAI’s launch guidance tells users to keep the capture focused and avoid secrets or sensitive material.[3] Realistic sample data is useful; passwords, tokens, private keys and unrelated confidential windows are not.[7]
Figure 1 | How Record & Replay turns a demonstration into a reusable Skill.
Test the Skill on a different case
Take a configured bug issue. The form is straightforward, but a team may also care about naming conventions, project assignment, required labels, priority rules and description structure. OpenAI lists a correctly configured issue among the examples for Record & Replay.[2]
The next issue will have different content. Its title, body or assignee may change, while the sequence of creating the issue, filling the required fields, submitting it and checking the result stays more stable. The team rules are less obvious: one class of bug may always carry a particular label, or production incidents may follow a different priority rule.
Before recording, state the goal and the inputs that are likely to vary. Use realistic sample data, complete one focused run, then stop. OpenAI’s guidance is to keep recordings short and complete rather than carrying on into unrelated cleanup.[7] Creating the issue and then clearing an inbox teaches two workflows when you only wanted one.
After Codex drafts the Skill, read it before replaying anything. A recording shows what happened, but one run may not explain why a choice was made. If the demonstration uses High priority, for example, Codex can see the value. It cannot infer from that observation alone whether every production bug should be High or whether this case happened to be severe.
OpenAI’s guide treats refinement as part of the workflow and calls out hidden preferences such as naming conventions, field defaults and decision points.[7] Remove sample values that accidentally became durable rules. Add the condition behind a branch. Write down the team convention another operator would need before making the same choice.
Then change the case. If the recording used:
Title: Checkout fails after coupon removal
Priority: High
Project: Paymentstest another issue with a different title, project or priority. Review the result, not just whether Codex reached Submit. Was the issue created? Are the required labels present? Did the priority follow the rule? Is the description still in the expected shape?[7]
Write down what counts as a correct result before trusting the replay. If you still have to correct the same fields by hand on the second case, edit the Skill and test again.
Some workflows are not worth recording
OpenAI describes Record & Replay as a fit for stable, repeatable work that is easier to show than describe.[2][3][4] A short written procedure may already be enough. In that case, write the Skill directly. Skills were designed to package reusable instructions, resources and scripts, and Codex can use them explicitly or select them for a matching task.[6]
For a one-off exploratory task, use a normal prompt. When each run needs a fresh plan, the interface changes constantly, or nobody can state what success looks like, there is little stable procedure to extract.
Figure 2 | When should you use Record & Replay?
Keep scheduling and packaging separate
Record & Replay still creates a Skill. Codex Automations can combine instructions with optional Skills and run them on a schedule, with completed work landing in a review queue.[6]
Plugin packaging becomes relevant when the workflow needs to be installed and shared as a broader bundle. By 26 March, OpenAI described Codex Plugins as installable workflow bundles that could package Skills, optional app integrations and MCP server configurations for reuse across projects or teams.[5]
References
- OpenAI, ChatGPT & Codex changelog, 18 June 2026 Record & Replay release. https://developers.openai.com/codex/changelog
- OpenAI, ChatGPT Release Notes, 18 June 2026 Record & Replay entry. https://help.openai.com/en/articles/6825453-chatgpt-release-notes-notes
- OpenAI, ChatGPT Business - Release Notes, 18 June 2026 Record & Replay entry. https://help.openai.com/en/articles/11391654
- OpenAI, ChatGPT Enterprise & Edu - Release Notes, 18 June 2026 Record & Replay entry. https://help.openai.com/en/articles/10128477
- OpenAI, ChatGPT Business - Release Notes, 26 March 2026 Plugins in Codex entry. https://help.openai.com/en/articles/11391654
- OpenAI, Introducing the Codex app, 2 February 2026; used for the established roles of Skills and Automations. https://openai.com/index/introducing-the-codex-app/
- OpenAI, Record & Replay, living documentation; used only to cross-check workflow mechanics, Skill draft structure, refinement and replay, not to backdate UI, plan or rollout changes after 3 July 2026. https://learn.chatgpt.com/codex/extend/record-and-replay