Anthropic Launches 'Teach Claude a Skill': Record Your Screen to Teach It Your Tasks

Imagine teaching your digital assistant a task the way you teach a new colleague: you record your screen as you perform it, explain your steps with your voice, and it learns it and re-executes it whenever you want. This is exactly what the "Teach Claude a skill" feature Anthropic launched offers. Instead of writing lengthy instructions or uploading explanatory documents, you demonstrate the task once, and Claude turns it into a reusable "skill." The idea seems simple, but it addresses what most exhausts AI tool users: repeating the same explanation every time.

How Does the Feature Work?

The mechanism is direct. From the plus (+) menu in the Claude desktop app, you select "Record a skill," then start a recording session during which you perform the task and explain your actions with your voice as you go. The system captures screen activity, mouse clicks, keystrokes, and voice commentary, then processes this demonstration into a structured, reusable skill definition. In practice, you do not describe the task, but perform it, and Claude learns the workflow by watching it in full rather than relying on a text description it might misinterpret.

Where Does It Appear, and For Whom?

The feature comes within the Claude Cowork interface in the desktop app, and is rolling out to Pro, Max, and Team plan subscribers. As a reminder, Cowork is the agentic work tool whose expansion to mobile and web we previously covered. "Record a skill" sits in the plus menu alongside existing options like adding files, browsing skills, connectors, and plugins. Team plans require at least five seats, while Max plans operate on rolling five-hour usage windows with no seat minimum.

The Broader Context: What Are "Skills" Anyway?

To understand this announcement's importance, it must be placed in the context of the "Skills" system Anthropic launched in October 2025. A skill is at its core a self-contained "task package": a folder containing instructions, scripts, examples, and resources, which Claude loads dynamically only when needed. These packages allow guiding Claude on a specific workflow: creating documents per your company's brand guide, analyzing data with your organization's methods, or automating recurring personal tasks. What's new today is that building a skill, which used to require writing an instructions file (SKILL.md) in a structured format, is now possible by simply recording and narrating — that is, it opened the feature to non-technical users.

Why Does This Actually Matter?

The real value is in overcoming the repetition barrier. Anyone who uses AI tools for recurring tasks (weekly reports, file organization, spreadsheets in a certain format) knows the cost of re-explaining the same task every session. Turning "demonstrate once" into a "permanent skill" moves the tool from an assistant that starts from scratch every time, to a tool that accumulates your experience with it. This aligns with the deeper 2026 trend we've tracked: the shift of AI assistants from momentary responses to automating the full workflow via organized tools and skills.

A Dimension Worth Attention: Privacy

Along with the feature's convenience, a point remains that calls for awareness. Screen recording, by its nature, captures everything visible during the session: clicks, keystrokes, and any sensitive data that may be visible. The practical rule is to record your skills in a clean environment free of private information (passwords, customer data, personal messages), and to review what the skill captured before saving or sharing it. The feature is powerful, but granting a tool the ability to "watch" your screen is a decision worth conscious control, like any expansive permission.

This launch remains an example of a deeper shift in how we interact with AI tools: from "telling" them what to do, to "showing" them how to do it. For the developer and content producer, the feature opens a practical door to automating recurring office tasks without writing complex instructions — provided one balances the convenience of automation with awareness of what it reveals to the tool. Try it first on a simple, non-sensitive task, and assess the accuracy of what Claude learned, before relying on it for a critical workflow. (Disclosure: this news concerns a feature from Anthropic, the maker of Claude, and is presented objectively without promotion.)