This is the same thing we asked clients to do whenever we needed to capture how they run a process: record your screen while you do the task, and narrate what you're doing as you go. The difference now is that instead of a person reviewing that video and designing a repeatable process, Claude does it. That's the whole idea behind learning to teach Claude a skill by recording your screen: you demonstrate the work once, and Claude turns the recording into a reusable skill you can call any time.

Running a business is all about systems, and skills already put building systems on easy mode. This new feature puts building the skills themselves on even easier mode.

First, what a skill actually is

A skill is plain-English instructions for AI, like a saved prompt. Claude, or any other model, follows those saved instructions the same way every time. But a skill can go beyond a single text file. It can be a whole bundle that includes templates, Python scripts, references, boilerplate, examples, design systems, and personas.

You write it once, so you never have to re-explain the process. "Here's how we format the weekly sales report." "Here's how we rename and file incoming invoices." "Here's how we turn a call transcript into a follow-up email and a task list." Write it down a single time, and Claude runs that same process, using the resources inside the skill, over and over.

What's often misunderstood is that a skill isn't always 100% processed by AI. You can build a skill sandwich, where AI is the bread and the exact, must-be-right work in the middle runs as real code, usually a Python script.

That matters because it puts the right tool on each part of the job. The steps that must be exact (a calculation, a data transform, pulling a number, validating a format) run as real code, so they're reliable. The steps that need flexibility (reading your request, interpreting the result, choosing the next move) run on AI. You get the support of AI with the reliability of code, stitched into one process, instead of hoping the model gets the math right every time.

Work is going "skills-first"

Most people still use AI conversationally. They open a chat, explain what they need, get a result, and next week open a new chat and explain the same thing all over again. It works, but you pay for a lot of repetitive work every time, and you're at the mercy of how well you described it that day.

Skills-first is a different habit. Instead of asking "how do I get AI to do this thing right now," you ask "what do I do over and over that I can turn into a skill?" You spend a little time building the skill once (most take under an hour), and then the task gets faster, more consistent, and easier to hand off, in perpetuity.

Landing-page creation, a recurring report, a proposal draft, a client onboarding packet, a weekly analysis, a file-naming convention: they're all perfect candidates for skills. The teams pulling ahead are the ones who stopped re-explaining the same task and started turning their best processes into repeatable assets.

How to teach Claude a skill by recording your screen

The "Record a skill" feature lets you build a skill by demonstration instead of description. You start a recording, do the task on your own screen the way you always do it, and talk through your reasoning as you work. Claude captures the whole thing: the screen activity, your clicks, your typing, and your voice. When you stop, it analyzes that sequence and turns it into a skill it can run again later.

The narration matters as much as the clicks. The recording shows Claude what you did; your voice explains why. "I always check this column against last month before I send it." "If this field is blank, skip that step." "We never publish without a second set of eyes." That reasoning is exactly the stuff that usually lives only in your head, and it's what separates a skill that actually works from one that mechanically repeats motions.

It's the same idea as grabbing a new hire, screen-sharing, and saying "here, let me show you how we do this." Except you only have to do it once, and the trainee never forgets.

Don't write it by hand

The entire skill bundle can be built by Claude, with your guidance. You never touch the code or the instruction files yourself. Whether you record a skill or just ask Claude to build one in a chat, Claude writes everything: the SKILL.md, any Python scripts, the templates, the examples, and then packages the whole bundle.

You don't need to know Markdown, you don't need to write a line of Python, and you don't need to understand how the pieces are wired together. Claude assembles it, then gives you a preview and a "Save Skill" button to activate it in your Cowork.

What you do need to bring is judgment, and that's the one thing you should not hand to the AI. You decide what a good output looks like, which steps are non-negotiable, where the checks belong, and what "done right" actually means for this task. Claude can build the skill; it can't tell you whether the skill produces the result your business actually needs.

Design the skill around the outputs you genuinely want, be specific about the rules and the exceptions, and review what it produces before you rely on it. Get that right, and you have a repeatable process you can trust. Skip it, and you've automated a vague idea of the task, which is often worse than doing it by hand.

What makes a recording worth keeping

Treat the first recording as a draft, not a finished product. A few things, drawn from how we've built skills the manual way, separate a recording that becomes a genuinely useful skill from one that just parrots your clicks:

Why this matters

Since AI hit the mainstream, the bottleneck has been translating what you know into something the model can reliably repeat. Half of what you do is muscle memory you've never put into words, and describing a process in the abstract is genuinely hard.

Recording a skill attacks that bottleneck directly by letting you demonstrate instead of describe. You don't have to be good at prompting. You don't have to think like an engineer. You just do the task the way you already do it, explain yourself the way you would to a colleague, and let Claude take it from there.

Knowledge that was previously trapped in someone's head becomes something the whole team can run. And it compounds. Do this across a dozen recurring tasks, and you've quietly built a library of your team's best practices that Claude can execute on demand.

It's also a sign of where this is heading. Anthropic keeps positioning Cowork as less of a chatbot and more of a digital teammate, and "learn by watching me once" is a very teammate-like move. It isn't unique to Claude, either: OpenAI added a comparable record-and-replay capability to its Codex tool, which tells you this "show, don't tell" approach to teaching AI is becoming a standard way of working, not a one-off.

One caveat worth repeating: keep a person in the loop. A recording captures what you showed it, including any mistakes or shortcuts. Review the skill it produces, test it on a real example, and keep a human check on anything that touches money, customers, or publishing. The review gate is a feature, not a limitation.

The takeaway

If you've been meaning to systematize the tasks you do over and over but never found the time to write it all down, the excuse just got a lot weaker. Show Claude while you execute the actual process you want it to repeat. It takes very little extra time beyond the work you're already doing, and what you get back is a process the whole team can run.