Codex Record & Replay: Show Your Workflow Once, Get a Reusable Skill

Every AI coding tool so far has asked the same question: describe what you want automated, in words, precisely enough that the model can reconstruct the steps. Codex Record & Replay, which shipped June 18, 2026, asks a different question entirely: what if you just showed it? You perform the workflow on your Mac once while Codex watches, and it produces a reusable, editable skill that replays the same logic against new inputs — without you writing a single line of prompt.

The timing is pointed. June 18 is also the day Gemini CLI's free tier went dark. While one workflow tool exited, a new automation paradigm arrived. Record & Replay is the first mainstream implementation of workflow capture for AI coding — and the output format it generates maps directly to the open SKILL.md standard that Claude Code, Cursor, and the skills-hub registry already speak.

The shift: show, don't describe

Prompt engineering for workflow automation has always had a gap. The workflows that are hardest to automate — filing an expense report against your company's specific form, creating a correctly configured Jira issue with the right project and label taxonomy, publishing a video through a multi-step CMS — are exactly the workflows hardest to describe in text. They have hidden defaults. They depend on which fields exist in your specific instance. They have context that a new employee would absorb by watching, not reading.

Record & Replay inverts the authoring model. Instead of writing a prompt that describes the workflow, you demonstrate the workflow. Codex observes your actions and the on-screen content, then produces a natural-language skill document that captures what you did, what decisions you made, and how to verify that the result is correct.

This matters beyond convenience. The resulting skill is inspectable and editable. You can read exactly what Codex understood from your demo, fix misinterpretations, add edge cases it didn't see, and share the result with your team — all before running it on real data.

How recording works

The capture flow is five steps. It lives in the Plugins panel, not the chat interface, which signals that OpenAI sees this as infrastructure rather than a conversational feature.

1. Open Plugins → select "Record a skill"
2. Give Codex context: name the workflow and describe what it accomplishes
3. Grant screen and input observation permission when prompted
4. Perform the workflow on your Mac at a natural pace
5. Stop recording via menu bar icon, the overlay button, or by saying "stop recording"

During capture, Codex is watching your actions and the window content visible at each step. It is specifically designed to learn theintentbehind each action, not just replay the pixel coordinates — so when you click a "Submit" button, it understands “submit a form” rather than “click at (782, 441).” This is what makes replay work when the layout changes slightly or the data differs.

Two things to avoid: secrets and sensitive data. Anything visible on-screen during recording can appear in the generated skill. If your workflow passes through a credentials page, either mask the screen or record a sanitized version with placeholder values.

What Codex generates

After recording, Codex produces a structured skill document covering four areas: when to use the workflow, what inputs it requires, the step-by-step procedure, and how to verify the result. This maps directly to the SKILL.md structure developers on skills-hub already know.

---
name: submit-expense-report
description: File an expense report in Concur for a single-day travel event.
  Handles project code selection, receipt upload, and manager routing.
version: 1.0.0
category: productivity
platforms:
  - CODEX
---

## When to use
When you have a same-day travel receipt to submit and the project code is known.
Not suitable for multi-day trips or non-travel expenses (use the general form).

## Required inputs
- EXPENSE_DATE: date of the expense (YYYY-MM-DD)
- AMOUNT: total amount in USD
- PROJECT_CODE: 6-digit project code from the finance system
- RECEIPT_PATH: local path to the receipt image or PDF

## Steps
1. Open Concur at https://[your-concur-domain]/expense
2. Click "New Expense Report" → select "Single Day Travel"
3. Set the expense date to EXPENSE_DATE
4. Enter AMOUNT in the Amount field; currency defaults to USD
5. Paste PROJECT_CODE into the Project field; wait for the dropdown to confirm
6. Upload RECEIPT_PATH via the attachment button
7. Click "Submit to Manager" — verify the confirmation banner appears

## Verification
The workflow is complete when the green "Submitted" banner appears and the
report appears in the "Pending Approval" queue under Reports > My Reports.

The generated skill is immediately editable. In practice, the most useful refinement is adding the hidden defaultsthat your demo used implicitly — the expense category you always choose, the manager routing rule that's specific to your team — so that teammates who replay it on their accounts don't need to guess.

Replay in practice

To run a captured skill against new data, start a new Codex thread and ask it to apply the skill by name, passing the variable values for that run. Codex resolves the variable placeholders and then dispatches the workflow using whatever tools are available — Computer Use for GUI interactions, browser actions for web steps, and plugins for anything with a structured API.

> Apply the submit-expense-report skill with:
>   EXPENSE_DATE = 2026-06-21
>   AMOUNT = 84.50
>   PROJECT_CODE = 881204
>   RECEIPT_PATH = ~/Downloads/hotel-receipt.pdf

The four-surface reach — Codex, Computer Use, browser, and plugins — is what makes this useful for real workflows. An expense report submission that involves navigating a legacy web form, uploading a local file, and confirming via a desktop notification can all be captured in one recording and replayed without splitting it across tools.

Variables are resolved by name, not position, which means you can store commonly-used values in Codex's context and reference them across multiple skill invocations in a session. A weekly report skill invoked every Monday can pull the date from context automatically if you set it once at session start.

Cross-tool portability

The skill document Codex generates is not locked to Codex. It uses the same frontmatter structure as the SKILL.md open standard, which means it can be registered, searched, and installed via skills-hub without modification.

# publish a Codex-generated skill to the registry
npx @skills-hub-ai/cli publish ./submit-expense-report.md

# install it in Claude Code
npx @skills-hub-ai/cli install submit-expense-report

# install it in Cursor (.mdc rules format)
npx @skills-hub-ai/cli install submit-expense-report --format cursor

The portability gap is on the execution side, not the description side. A Codex-generated skill that includes Computer Use steps will not replay those steps inside Claude Code — Claude Code doesn't have Computer Use. But the skill document still serves as a precise procedural checklist that a Claude Code session can execute via terminal commands and file writes where possible, and flag for human completion where it can't.

For teams that use multiple tools, this creates a natural handoff pattern: capture the initial workflow with Codex's Record & Replay, publish to skills-hub, and let each team member install it into whichever tool they use. The skill document travels; the execution adapts.

What to capture first

Not every workflow benefits equally from recording. The highest-return candidates share three properties: they're recurring (you do them weekly or more), they involve more than two applications, and the steps are stable enough that you can demonstrate them identically twice. Here are the six categories with the strongest ROI.

1. Finance ops       — expense reports, invoice approvals, PO creation
2. Dev ops           — deployment checklists, staging smoke tests, hotfix filing
3. Content ops       — video publishing, newsletter scheduling, social card export
4. HR + admin        — time-off requests, onboarding ticket creation, badge access
5. Reporting         — weekly metric pulls, dashboard screenshot + Slack delivery
6. CI triage         — test failure categorization, flake marking, re-run sequencing

Workflows to avoid for your first recording: anything involving live credentials or secrets visible on screen, multi-hour workflows with long wait states, or workflows where each run takes a meaningfully different path (branching logic is better handled with a written skill than a recorded one).

For dev workflows specifically, the best recordings are the ones your team currently documents in a Confluence page as a numbered checklist. If it's checklist-shaped and multi-app, it's a good recording candidate. The Codex-generated skill then becomes the living version of that doc — one that can actually run.

Current limits and geography

Record & Replay launched with four constraints worth noting before you build a team workflow around it.

Platform:    macOS only (Windows not yet supported)
Geography:   Excluded at launch: EEA, United Kingdom, Switzerland
Plan:        Requires ChatGPT Plus, Pro, Business, Enterprise, or Edu
Dependency:  Computer Use must be enabled (check Settings → Features → Computer Use)

Enterprise:  If your org uses requirements.toml, add:
             [features]
             computer_use = true
             to enable Record & Replay for managed accounts

The macOS restriction is tied to Computer Use's system integration, not an architectural choice — OpenAI has indicated Windows support is on the roadmap. The EEA/UK/Switzerland exclusion is a regulatory timing issue, not a permanent restriction.

For teams with mixed operating systems, the cross-platform portability strategy above still holds: record on a Mac, publish the skill to the registry, and let Windows teammates install and execute the non-Computer-Use portions via their own tools.

The most interesting thing about Record & Replay is not the feature itself — it's what it signals about the trajectory of AI workflow tools. The SKILL.md standard was designed for human-authored instructions. With Record & Replay, AI-generated skills become a first-class input. The registry now accepts skills from three sources: written by developers, synced from official repositories, and demonstrated on a Mac. Browse what's already there at /browse?category=productivity or publish your first captured skill with npx @skills-hub-ai/cli publish ./your-skill.md.