How to Teach AI a Task by Recording Your Screen: ChatGPT Record & Replay vs Claude Record a Skill (2026 Guide)

You know the feeling. You spend ten minutes carefully writing a prompt that explains every click, every folder name, every “if the amount is over 500 then ask me” rule. The AI still misses the preference you never quite put into words. Then next week you do it all again.

There is a better way in 2026: teach AI by recording your screen. You perform the task once, talk through the decisions if you want, and the system turns that demonstration into a reusable skill. ChatGPT calls it Record & Replay (in Work mode or Codex). Claude calls it Record a skill (in Cowork). Both live in the desktop app on a Mac. Show once, automate forever — or at least until the UI changes.

This guide walks you through exactly how both features work, when they are worth using, the precise steps, ready-to-copy prompts, and the real limitations. Primary keyword “teach AI by recording your screen” appears early and naturally; secondary terms include Record & Replay, Claude Cowork Record a skill, and Computer Use.

Why Record-and-Replay Beats Writing Long Prompts

Long prompts fail for the same reason onboarding documents fail. People leave out the tacit knowledge — the exact column order they always use, the way they skip a field when it is blank, the naming convention that only makes sense once you have seen it. Demonstration captures that tacit layer.

Official OpenAI documentation states that Record & Replay is ideal “when the workflow is repetitive, depends on your preferences, or is easier to show than to describe in a prompt.” Anthropic’s help center says the same for Claude: instead of writing a skill by hand, you record yourself doing the task and let Claude build the skill from what it observes.

Real-world examples that fit well:

• Filing an expense report (upload receipt, choose category, fill description, submit).
• Downloading a recurring report from an internal dashboard and saving it with a consistent filename.
• Creating a correctly configured issue or ticket with your preferred labels and template.
• Publishing a video or social post that pulls metadata from a spreadsheet.
• Renaming and organizing a batch of files according to a house style.

A good first task has three qualities: the steps stay roughly the same every time, the success criteria are clear, and you do it often enough that the time saved matters. If the task requires heavy judgment that cannot be narrated, or if a proper connector or plugin already exists, stick with the connector.

One creator who tested Claude’s version reported cutting multi-hour research workflows down to about 30 minutes after the skill was refined — with the usual caveats about usage limits and occasional fragility.

Think of it as the difference between writing a recipe and cooking the dish while someone watches. The recipe can list ingredients and times, but the demonstration shows the exact heat, the moment you taste and adjust, and the way you plate it. Both ChatGPT and Claude are trying to turn that cooking session into a recipe the AI can follow later. The quality of the skill depends heavily on how clearly you demonstrate and how carefully you review the draft.

Teams that process the same type of form or report every day see the biggest gains. Individual knowledge workers who handle a recurring personal workflow (expense reports, client onboarding checklists, weekly content packaging) also benefit quickly. The first recording takes a few minutes; the second and third runs often take almost none.

What You Need Before You Start

Both features currently require the desktop application on macOS. Neither recording flow works on Windows yet (as of late October 2026). Computer Use itself is available on Windows for ChatGPT, but the “record a skill by watching me” path is Mac-only.

For ChatGPT Record & Replay:

• Latest ChatGPT desktop app on Mac.
• Work mode enabled (switcher) or Codex.
• Computer Use plugin installed and enabled. On macOS this needs Screen Recording and Accessibility permissions.
• A paid plan is typically required for full Computer Use and skill features; exact entitlements can vary — check your account.
• Regional note: early launches excluded the EEA, UK, and Switzerland for some related features; confirm current availability.

For Claude Record a skill:

• Latest Claude desktop app on Mac.
• Cowork workspace (not regular chat).
• Pro, Max, or Team plan (not Free or Enterprise according to official help docs).
• Screen recording and Accessibility permissions granted the first time.

Before recording, pick a clean example. Use dummy receipts or sample files rather than real sensitive data the first time. Close unnecessary windows so the AI is not distracted by background noise.

How to Record a Skill in ChatGPT (Record & Replay)

OpenAI’s official docs (learn.chatgpt.com/docs/extend/record-and-replay) give this flow:

1. In the ChatGPT desktop app, select ChatGPT and turn on Work in the switcher, or switch to Codex. Open Plugins.
2. Open the + menu.
3. Select Record a skill.
4. Review the suggested prompt. Add helpful context such as the goal, what changes each run (date range, file names, amounts), and any preferences. Submit.
5. When the chat asks for permission to record your actions, approve once you are ready.
6. Perform the workflow on your Mac exactly as you normally would. Keep the recording focused on the task.
7. Stop recording from the menu bar, the on-screen overlay, or by telling the chat you are done.

During recording, ChatGPT observes the actions and window content needed to learn the workflow. After you stop, it inspects the capture and drafts a skill. The skill describes when to use the workflow, what inputs it needs, the steps, and how to verify the result. You can ask for refinements.

To replay: start a new ChatGPT or Codex chat and ask it to use the generated skill. Supply the new inputs (different receipts, different report period, etc.).

Example goal prompt you can adapt before recording:

I will show you how I file an expense. Watch where I put the receipt, how I name the expense, which category I choose, and which fields I fill. After the recording, turn this into a reusable skill. Use only sample data with no real account numbers.

After the skill is created, a useful verification line is: “When you’re done, tell me how many receipts you found and how many you logged.”

Some walkthroughs mention installing a Record plugin and typing @record, but the official documentation centers on the Plugins → + → Record a skill path. Stick to the documented steps if the shortcut does not appear.

After the skill is drafted, spend time in the review step. Ask ChatGPT to show you the steps it extracted. Correct any misread field names or incorrect assumptions. You can also request that the skill always pause before submitting or sending anything. This review is where most of the reliability is created. Many users treat the first generated skill as a draft and refine it two or three times before trusting it on real data.

If you use Codex as well as ChatGPT Work, the same Record & Replay capability is available there. The skill can then be reused in coding or multi-step agent sessions that combine computer use with other tools.

How to Record a Skill in Claude (Cowork)

Claude’s official help center (support.claude.com) describes the process:

1. Update to the latest Claude for Mac and open Cowork.
2. Start recording one of two ways: click the + button in the composer and choose “Record a skill,” or go to Customize → Skills → Add → Record your screen.
3. Grant Accessibility (for mouse/keyboard) and Screen recording permissions if prompted. macOS may ask you to restart Claude.
4. Select your microphone and start recording.
5. Do the task while talking through it. Narration is important — explain why you skip a step or how you choose between options. Claude uses both the screen and your voice.
6. Click Done when finished. Recording length is roughly 10 minutes with a countdown near the end.

Claude then starts a Cowork task, reviews the recording, and proposes a skill for you to review before saving. You can edit it. Once saved, the skill can be invoked by name or in future Cowork sessions.

Example narration style that works: “I’m opening the monthly report, renaming it with the project code and today’s date, then moving it into the client folder. If the total is under 100 I skip the approval flag.”

Because Claude explicitly listens to your voice, the quality of narration matters more here than in the ChatGPT flow. Speak clearly and at a normal pace. Mention the “why” behind choices rather than only describing the clicks. After the proposal appears, read the skill description carefully. Claude sometimes includes extra setup steps that were part of your recording but not essential to the core task; you can ask it to remove those.

On Team plans you can later share the skill with colleagues. Shared skills become especially valuable when several people perform the same standardized process (invoice entry, ticket triage, report packaging).

Quick Comparison

Best Practices for Reliable Skills

A skill is only as good as the demonstration and the review.

• Narrate decisions. Especially in Claude. Say the exceptions out loud: “If the date is in the future I leave the status as draft.”
• Keep the recording tight. Do only the task you want automated. Extra browsing confuses the model.
• Use sample data first. Protect sensitive information.
• Review and edit the draft skill. Both systems produce an inspectable skill (often conceptually similar to a SKILL.md). Fix incorrect assumptions before you rely on it.
• Test with variations. Run the skill on a slightly different set of inputs and watch the first few executions.
• Add verification. End the skill prompt with a clear success check (count of items processed, final file location, confirmation message).
• Manage permissions carefully. Screen recording and accessibility are powerful. Grant them only while recording and review which apps Computer Use can touch.

Ready-to-copy verification prompt after the skill is created:

Use the expense-filing skill on these three sample receipts. When finished, list: (1) how many receipts processed, (2) the final category chosen for each, (3) any fields left blank, and (4) the confirmation or error messages. Do not submit anything real.

Limitations, Common Pitfalls, and When Not to Use It

These are demonstration tools, not full robotic process automation replacements.

• Mac desktop only for the recording path as of October 2026. Windows users can use Computer Use for live tasks but cannot yet record skills this way.
• Plan and region restrictions apply. Claude requires Pro/Max/Team; ChatGPT features tied to Computer Use have plan and regional caveats.
• UI fragility. If a button moves or a site redesigns, the skill can break. Plan to re-record or edit when interfaces change.
• Not encrypted vaults. Recording captures what is on screen. Do not demonstrate with production credentials or highly sensitive data.
• Usage and speed. Complex skills can consume significant context and run slower than a well-written prompt or native connector.
• Judgment-heavy tasks that cannot be shown clearly are better left to normal prompting or human review.

If a proper integration or plugin already exists for the app (Gmail, Slack, your CRM, etc.), prefer that over screen recording. Recording shines for desktop apps and web UIs that have no connector.

Privacy and security deserve extra attention. Both features require screen recording permissions. Anything visible on your screen during the recording can be observed by the model. Avoid recording while sensitive passwords, personal financial data, or confidential client information is visible. Use a clean desktop profile or a dedicated workspace if possible. After creating the skill, review exactly which actions it is allowed to take and set approval modes where available so the AI pauses before irreversible steps such as sending emails or submitting forms.

Common troubleshooting items:

• The “Record a skill” option does not appear — update the desktop app, confirm you are in Work/Cowork, and check that your plan includes the feature.
• Permissions errors — grant Screen Recording and Accessibility in macOS System Settings, then restart the app.
• Skill misses a step — re-record with clearer narration or edit the generated skill text directly.
• Skill works once then fails — the target UI may have changed; re-record the affected portion.

What’s Next / How to Extend

Once you have a reliable skill:

• Combine it with scheduled tasks in ChatGPT or Claude so the workflow runs on a cadence (daily report, weekly expense check).
• In ChatGPT, the skill can be replayed with Computer Use, browser actions, or connected plugins.
• In Claude, skills can be shared on Team plans and installed from the directory.
• Keep a short library of your highest-value skills and review them quarterly.

Try these low-risk first tasks: file three dummy expenses, download and rename a sample report, or create a test ticket with your preferred labels.

Conclusion

Teaching AI by recording your screen is one of the most practical advances in everyday AI workflows in 2026. ChatGPT’s Record & Replay and Claude’s Record a skill both let you demonstrate a stable, repetitive task once and turn it into an editable, reusable skill. The features are currently Mac desktop only, work best on tasks with clear success criteria, and still need human review — but they eliminate a large class of “I have to re-explain this every time” friction.

Start with one small, frequent task this week. Record it carefully, review the generated skill, test it on sample data, and only then let it run on real work. The time you save on the second and third runs is usually worth the five minutes of setup.

Have a workflow that finally feels automatable? Open the desktop app, hit Record a skill, and show it once.

Last verified: October 25, 2026. Features are new and changing — always confirm the latest steps in the official apps and docs.