Claude Can Now Watch You Work. Then It Never Asks Again.

You already trained a new hire this way. Now you can train Claude the same way, and it never forgets a step.

It is 8:40 on a Monday morning in Manchester. Priya runs a five-person homeware shop, and before she opens the till, she does the same ten minutes of nothing.

Three supplier websites. Three logins. Three product grids that have not changed their design since 2014. She checks what is low, keys the same six items into the same order form, and closes the tab.

None of these suppliers have an app. None of them have an API. Nobody is ever going to build her a Zapier integration for a wholesale ordering portal that still uses a table layout from the last decade.

Until this month, that meant Priya was stuck doing it herself, forever. Anthropic's newest Claude feature says otherwise.

Every Monday, Priya Does The Same Ten Minutes Of Nothing

Here is how it works right now, for a lot of small business owners:

• You do a repetitive task inside a website or piece of software that has no official integration.
• You cannot connect Claude, ChatGPT, or any agent to it directly, because there is no API for the agent to call.
• So you keep doing it by hand, every single week, because teaching someone else to do it exactly right takes almost as long as just doing it yourself.

That last point is the real cost. Not the ten minutes. The fact that the task lives entirely in your head, and nobody, human or AI, can take it off your plate without you explaining it in full, every time.

You are the bottleneck. Not because you are slow. Because you are the only documentation the task has ever had.

What Is Claude's "Record a Skill" Feature?

Record a Skill is a new tool inside Claude Cowork, Anthropic's agentic workspace, that lets you teach Claude a task by recording your screen once instead of writing instructions. You click the "+" button in Cowork, select "Record a skill," then do the task the normal way while you talk through it out loud. Claude watches your clicks, your typing, and listens to your narration, then drafts a skill it can run again by itself.

Anthropic announced it on 21 July 2026, describing it plainly: "Record your screen while you do a task, talk through it as you go, and Claude turns it into a skill it can run again." It is rolling out now to Claude Pro, Max, and Team subscribers inside the Claude for Mac desktop app, under Cowork.

Here is the micro version of what that looks like in practice. YouTuber Brock Mesarich, who tested it live for his channel AI for Non Techies, recorded himself opening YouTube Studio, filtering his comments, and copying the useful ones into a spreadsheet to flag video ideas. He talked through his reasoning as he clicked. When he hit "Done," Claude had turned 73 screen actions into a working skill.md file, one that correctly picked up a detail he never even said out loud: never touch, like, or delete the original comments. It inferred that rule from what he did not do.

That is the difference between a prompt and a demonstration. A prompt tells Claude what you want. A demonstration shows Claude how you actually think while you do it, including the judgment calls you never bothered to write down because they felt too obvious to mention.

Why This Is Not Just Another Way To Say "Chatbot"

If you have read anything about Claude this year, you have read about Model Context Protocol, MCP for short, the open standard that lets Claude connect directly to apps like Slack, Google Drive, and thousands of others. MCP just shipped its biggest spec update yet on 28 July 2026, and Claude's connector directory now lists more than 950 servers used by millions of people daily.

MCP is brilliant for the tools that bother to build a connector, and if you missed what changed in MCP's biggest update yet, it is worth a read. It is useless for the ones that do not. Priya's supplier portals will never get one. Neither will the council licensing site your accountant logs into, the niche booking system your clinic has used since 2016, or the ancient stock system your competitor's warehouse still runs on.

Record a Skill is Anthropic's answer to that specific gap. Not a replacement for MCP or a tool like Zapier, which already connects Claude-style agents to more than 9,000 apps with proper APIs. A patch for everything MCP and Zapier cannot reach, because a screen recording does not need the software on the other end to cooperate at all. It just needs to render on a screen and respond to clicks, the same way it already does for you.

Is This Safe For A Business With Sensitive Data?

This is the honest objection, and it deserves a straight answer instead of a marketing one.

Anthropic's own warning is blunt: do not type passwords or secrets, and do not display sensitive information or private conversations while recording, because everything on screen is captured for the length of the session, along with anything you say. The video and audio themselves are not kept after Claude processes them, but a set of screenshots from the session is retained as part of the skill.

The practical fix is simple. Record the boring, structural part of a task, the clicking and the page-hopping, and keep genuinely sensitive steps, like entering a client's banking details, outside the recording. Most repetitive small business tasks, restocking a catalogue, filing a routine report, updating a tracker, do not touch anything you would not show a new employee on their first day anyway.

How To Record Your First Skill

Here is the exact sequence, based on Anthropic's published steps and how it plays out for a real workflow like Priya's supplier restock:

1. Open Cowork in Claude for Mac. Record a Skill currently lives only inside Cowork on the Mac desktop app, on Pro, Max, or Team plans. It is not yet available in regular chat, on Windows, or on Free and Enterprise tiers.
2. Grant screen and accessibility permissions. macOS will ask for these the first time. Claude needs them to see your screen and track your clicks.
3. Click the "+" button, then "Record a skill." Or go to Customize, then Skills, then "Add," then "Record your screen."
4. Start recording and do the task normally, out loud. For Priya, that means logging into the first supplier site, checking the low-stock items against her own tracker, and placing the order, while she says why she is choosing each quantity, not just what she is clicking.
5. Click "Done." Claude opens a Cowork task, reviews the screenshots and narration, and drafts a skill, usually within a couple of minutes for a short recording.
6. Review before you save. Claude proposes a new skill or an update to an existing one, and shows you the skill file it wrote. Read it. Fix anything it misunderstood. Only then click "Save."
7. Run it, and watch what it caught that you did not say. Good demonstrations often surface rules you follow automatically without noticing, the same way Mesarich's skill knew not to touch the original comments.

SEO consultant Eliot Prince used the same feature to rebuild a monthly client-reporting workflow he says used to cost him a full day: logging into Google Search Console, pulling the numbers, writing commentary, and drafting the client email. He recorded himself doing it once, let Claude turn it into a skill, corrected the branding, and then set it to run on a schedule. What came out the other side, in his words, was a branded report and a drafted email, ready to send, without him touching a keyboard.

What You Have After This

A week from now, you have one task off your desk permanently, because you finally taught it instead of just doing it again.

A month from now, you have a small library of recorded skills for the handful of tasks that eat your Mondays: the supplier check, the weekly report, the tracker update, each one demonstrated once and never re-explained.

Six months from now, the tasks that used to require you personally, the ones with no API and no shortcut, are the same tasks a new hire could pick up in a day, because the skill file is the training manual you never had time to write.

Same shop. Same suppliers. A completely different Monday morning.

If you want to know exactly which of your own Monday-morning tasks are worth recording first, that is the first thing we map out in the 14-Day AI Readiness Sprint, before we touch a single tool.