Record the work once and repeat it with Claude.

Many tasks repeat not because they are intellectually difficult, but because someone has to remember a sequence of screens, decisions, and exceptions.

A demonstration preserves two layers: mechanics - which screen and button to use - and judgment - what you are looking for and how you know the work is complete.

A good candidate happens often, has a clear beginning and end, uses tools you can already open, and produces something another person can review.

The Claude/Cowork modal in the video is labeled Record a skill. It explains that the screen, clicks, typing, and voice are captured, then sent to Claude so it can propose a repeatable skill.

In Claude, skills are folders of instructions, scripts, and resources that load dynamically when they are relevant. That layer also exists outside recording and requires code execution to be enabled.

The equivalent feature in ChatGPT/Codex is called Record & Replay. OpenAI documents it in the ChatGPT desktop app for macOS while using Codex: demonstrate a workflow once and turn it into a reusable skill. Do not confuse it with ChatGPT Record, which records conversations and voice notes. The documented path is ChatGPT/Codex > Plugins > + > Record a skill, and Computer Use must be available and enabled.

The quality of the result is decided before you press Start recording. Choose a workflow, not an intention.

Weak task: “organize my operations.” Strong task: “every Friday, open the dashboard, filter the previous seven days, identify the five campaigns with the highest CPA, export the table, and save it in a folder.”

Before recording:

• close private conversations, passwords, tokens, and unrelated documents;
• prepare a view with non-sensitive data;
• open only the applications the task needs;
• define where the expected result should end up;
• decide what to do when a file is missing, a date changes, or an error appears.

The documented route in Claude for Mac is:

1. Open Cowork in Claude for Mac.
2. In the composer, select + > Record a skill. You can also use Customize > Skills > Add > Record your screen.
3. Read the warning and press Start recording.
4. Complete the task normally. Keep the microphone on if you will narrate.
5. Press Done when complete. If the demonstration drifted, choose Discard.

The capture bar shows that recording is active and counts the steps. Anthropic says a recording can last about ten minutes and displays a countdown when roughly one minute remains.

I am going to demonstrate a repetitive task called: [WORKFLOW NAME].
The final result must be: [VERIFIABLE OUTPUT].
The task begins when: [START CONDITION].
The task ends when: [END CONDITION].
These are the sources it may use: [APPS, FOLDERS, OR VIEWS].
These are the things it must not touch: [DATA, ACTIONS, OR ACCOUNTS].

Voice keeps the result from becoming a brittle choreography of clicks. Anthropic recommends speaking while you work because narration explains why you skip a step or choose one option over another.

Use these five phrases whenever a decision appears:

• “I am looking for...” names the signal.
• “I consider it valid when...” defines the criterion.
• “In this case I choose... because...” explains the preference.
• “If ... appears, then...” records an exception.
• “Do not...” marks a boundary.

I am looking for rows that meet [CRITERION], not merely rows with [SIMILAR SIGNAL].
I consider it valid when [OBSERVABLE CONDITION].
In this case I choose [OPTION] because [REASON].
If [EXCEPTION] appears, stop and flag it for human review.
Do not take [RISKY ACTION] without showing me a preview first.

Include at least one safe exception: what happens when there are no results, a date is empty, or a button does not appear. You do not need to create a real failure; describe the fallback.

After you press Done, Claude starts a Cowork task and reviews the recording. It may propose a new skill labeled NEW or an update to an existing skill.

Before choosing Save or Update, inspect scope, trigger, boundaries, and output. If the demonstration drifted, Dismiss is a good decision: the next version can be more focused.

Anthropic says video and audio are not retained after review. Screenshots from the session remain inside the Cowork task and are removed when that task is deleted.

Recording gives you a draft. Complete these fields so the skill still works next week:

name: weekly-campaign-report
description: Prepare the weekly campaign report from the previous 7 complete days.
Input:

Connected advertising dashboard.
Range: previous 7 complete days.
Folder: reports/campaigns/
Procedure:

Confirm the account and date range.
Filter and sort CPA from highest to lowest.
Export the five campaigns with the worst CPA.
Show a preview before sending.
Boundaries:

Do not change budgets.
Do not pause campaigns.
Stop if the account or date range is unclear.
Output:

One CSV and a five-line summary.

Official skill guidance recommends keeping each skill specific and repeatable, writing clear instructions, including examples when useful, and defining when it should run. Recording needs the same things: narrow scope and visible judgment.

Claude loads skills dynamically when they become relevant. To improve activation, describe the result and mention the skill name.

Use the [SKILL NAME] skill to complete [OUTCOME].
Work only with [ALLOWED SOURCES].
First confirm the account, date range, and output folder.
Return a preview before sending, sharing, deleting, or modifying anything.
If a source is missing or a decision is not covered, stop and ask me.

Anthropic confirms that scheduled Cowork tasks can use installed skills and plugins. They are available on paid plans and run remotely. A scheduled task cannot depend on a folder on your computer; if the workflow needs local files or apps, leave a manual step or use a compatible remote environment.

Every [DAY AND TIME], run the [NAME] skill to prepare [OUTPUT].
Use only [SOURCES]. Save the result in [ALLOWED LOCATION].
Do not send, publish, delete, or modify data without showing me a preview.
If a source is unavailable or data is missing, mark the task pending instead of guessing.

While recording, everything visible on screen and everything you say becomes part of the capture sent to Claude to build the proposal. Treat it as shared information, not as a private video that disappears without consequence.

• Recording captures what is visible and spoken; it cannot infer intentions you never explain.
• A change in layout, account, permission, or field name can break the recorded route.
• Skills can contain scripts or resources and should be reviewed before enabling, especially when they come from someone else.
• Actions with real consequences need a human review point.
• Scheduling uses the task's context and capabilities; never assume it can access the same local files.

Sources consulted

• Anthropic - How to create custom skills
• Anthropic - What are skills?
• Anthropic - Schedule recurring tasks in Claude Cowork
• Anthropic - Get started with Claude Cowork
• OpenAI - Record & Replay

Research reviewed August 1, 2026. The video confirms the Claude/Cowork name and visual flow. OpenAI's first-party source confirms the equivalent Record & Replay feature for ChatGPT/Codex on macOS. Account, region, version, and availability should still be checked in the app.