Claude can now watch you work and turn it into a skill it never forgets

Picture this: you finish uploading a new YouTube video and, out of habit, you run through the same mental checklist you always do. Did the end screen get added? Are the cards linked? Is there a playlist selected? Did you double-check the transcript for errors? You do this every single time, and it takes ten minutes you'll never get back.

Now imagine just doing that checklist once, out loud, in front of your screen — and never having to think about it again, because an AI watched you do it and can now do it for you.

That's not a hypothetical anymore. Anthropic just shipped a feature inside Claude that does exactly this, and it might be one of the most practical AI updates released this year — not because it's flashy, but because it solves a problem almost everyone runs into: teaching AI how you specifically do things, without spending hours writing instructions.

What is Claude's Record a Skill feature?

Record a skill is a new way to build a Claude Skill by showing Claude a task instead of writing instructions for it — you record your screen while completing a workflow, narrate your reasoning out loud, and Claude Cowork turns that demonstration into a reusable skill it can run again later. Instead of typing out a prompt describing every step of a process, you just do the task once while explaining your thinking, and Claude figures out the rest.

Anthropic's own Claude Help Center article on creating custom skills lays out exactly what's happening under the hood: instead of writing a skill by hand, you can record yourself doing a task and let Claude build the skill from what it observes — you send Claude a video of your screen, clicks, typing, and voice, and Claude proposes a skill for you to review before you save it.

This matters because until now, building a Claude Skill meant sitting down and writing a structured document. Building a custom skill meant authoring a SKILL.md file yourself — a structured markdown document with metadata, step-by-step instructions, and optional executable scripts, which gives you control and transparency, but takes writing, and the structure can feel intimidating to non-technical users. Record a Skill flips that entirely. Record a skill inverts it: the skill is authored for you from a demonstration.

Why does narration matter so much?

Here's the part people miss when they first try this: the value isn't just in the screen recording. It's in talking through your decisions while you work. Narration gives Claude context it can't get from your screen alone, like why you skip a step or how you choose between two options.

Think about it — a screen recording alone shows Claude what you clicked, but not why. If you skip adding a playlist because the video is a one-off tutorial but always add one for series content, that logic lives in your head, not on your screen. Saying it out loud is what turns a simple click-sequence into an actual reusable rule. The narration matters as much as the clicks: it's how Claude learns the reasoning behind each step, not just the sequence.

How do you actually set up a recorded skill?

You'll need access to Claude Cowork, which currently lives inside the desktop app. Recording a skill is available on Pro, Max, and Team plans, in Cowork in Claude for Mac — it isn't available in chat, on Windows, or on Free and Enterprise plans. Head over to claude.ai to check your plan, and if you're on a supported tier, here's the process:

1. Open the Claude desktop app and launch Cowork. This is a separate workspace mode from the regular chat interface.
2. Click the plus (+) menu and select "Record a skill." This sits alongside other options like adding files, connectors, and plugins.
3. Start recording and perform the task exactly as you normally would. Don't perform a "cleaned up" version — do it the real way, mistakes and all, since that's what teaches Claude your actual process.
4. Narrate your decisions the entire time. Explain what you're checking for, why you're clicking a particular button, and what would make you do something differently.
5. Stop recording and let Claude process it. Claude starts a Cowork task and reviews the recording, then proposes a skill.
6. Review the proposed skill before saving it. Anthropic's documentation notes you'll typically see one of two outcomes — a brand-new skill or an update to an existing one, depending on what Claude finds during review.

One safety note worth taking seriously before you start recording anything: Claude asks permission to record your screen and watch what you do — read the warning first, and do not type passwords or secrets, or show private information during the recording, since the whole point is that Claude is watching and processing everything on screen.

What's the difference between this and building a skill manually?

If you've used Anthropic's Agent Skills system before, you already know the traditional route. Anthropic provides pre-built Agent Skills for common document tasks like PowerPoint, Excel, Word, and PDF, and you can create your own custom Skills — both work the same way: once a Skill is available in your environment, Claude uses it automatically when relevant to your request.

Custom skills built the traditional way are genuinely powerful — Anthropic's open-source skills repository on GitHub shows just how far you can take them. Skills are folders of instructions, scripts, and resources that Claude loads dynamically to improve performance on specialized tasks, teaching Claude how to complete specific tasks in a repeatable way, whether that's creating documents with your company's brand guidelines, analyzing data using your organization's specific workflows, or automating personal tasks.

The catch has always been the authoring process. Writing a good SKILL.md file requires you to think like a technical writer — breaking down implicit knowledge into explicit, unambiguous steps. Record a Skill removes that bottleneck entirely for a huge chunk of use cases, especially ones that are visual and repetitive.

Which tasks are the best candidates for this?

Not every workflow benefits equally from being recorded instead of written out. The clearest guidance comes down to a simple filter: a recording shows what happened once, but a useful Skill must generalize beyond that one file, customer, date, or project — so test it with different inputs and inspect whether Claude learned the underlying rule rather than copying accidental details from the demonstration.

That means the best candidates are workflows you do often, that involve visible on-screen steps, and that follow a consistent internal logic even when the specific details change each time. A YouTube publishing checklist is a perfect example — the checklist items stay the same even though the video content changes every time. Same goes for things like weekly report formatting, expense categorization, or reviewing new leads against the same criteria.

Do you need to test the skill before trusting it?

Yes — and this is the step people are most tempted to skip. Review and test the Skill on a fresh example before relying on it, since the final two steps are essential. Run it once on a real, current task and actually check the output line by line. Did it apply the same logic you narrated, or did it just repeat the specific example you recorded? That distinction determines whether you've built something reusable or something that only works once.

It's also worth noting this is still a fresh feature. Availability is described as a phased rollout rather than a simultaneous flip for all eligible users, with several people reporting not yet having access shortly after launch. If the option isn't showing up for you yet, that's likely why — not a setup mistake on your end.

What should you actually record first?

Start small. Don't try to teach Claude your entire content pipeline or your whole client onboarding process in one recording. Pick one narrow, repeatable checklist — something you already do on autopilot — and record just that. A publishing checklist, a weekly report format, an inbox triage routine. Get comfortable with how the skill behaves, test it on a fresh example, and only then start stacking more of your workflow into Claude's memory.

The real shift here isn't the recording feature itself. It's the idea that teaching an AI how you work no longer requires you to become a prompt engineer or a technical writer. You just have to do your job the way you already do it — and say out loud why you're doing it that way. Everything else is Claude's problem to figure out.