Teaching AI how you like things done isn't as straightforward as it sounds. You can explain the same task over and over, save detailed prompts, or upload documents full of instructions, yet there's still a good chance the AI will miss a step or do something differently the next time you ask. It's a little like training a new coworker who keeps forgetting the workflow.

That's exactly the problem Anthropic is trying to solve with a new Claude Cowork feature called Teach Claude a skill. Instead of typing out long instructions, you simply record your screen while completing a task yourself and explain what you're doing as you go. Claude watches the entire process, learns it, and saves it as a reusable skill that it can perform again later.

How the feature works

Imagine showing Claude how you organize invoices in a spreadsheet, rename dozens of image files, prepare a weekly sales report, or update a project tracker. Once you've demonstrated the process, you don't have to explain every step again. Instead, Claude can repeat the workflow whenever you need it. That's far more useful than asking it to 'do it like last time' and hoping it remembers what you meant.

The process is straightforward: open Claude on your laptop, head into Cowork mode, click the + button in the chat box, and choose 'Record a Skill' under the Add files or photos menu. From there, you start recording your screen and walk Claude through the task. The AI captures both the visual actions and the verbal narration, building a contextual understanding of the workflow. It can then reproduce the same sequence of actions, adapting to minor variations if needed.

This shift from prompt-based to demonstration-based learning represents a significant evolution in how users interact with AI. Rather than translating every workflow into precise instructions, users can now simply show the AI what to do. This reduces cognitive load and makes advanced AI capabilities accessible to non-technical users.

Background and context

The concept of teaching AI by demonstration has roots in imitation learning and behavioral cloning, techniques used in robotics and autonomous systems. However, applying this to a general-purpose language model like Claude is novel. Anthropic has been steadily positioning Claude Cowork as a digital teammate, and this update pushes that idea further. The feature is now rolling out to Claude Pro, Max, and Team subscribers.

Claude Cowork was launched earlier this year as a mode that allows the AI to take over the user's mouse and keyboard to perform tasks on their computer. Early iterations were limited to responding to natural language commands, which often required multiple iterations to get right. The new 'Teach Claude a skill' feature addresses that limitation by enabling one-shot learning through demonstration.

Anthropic has not disclosed the exact technical implementation, but it likely involves the AI parsing screen recordings into a sequence of operations, mapping those to its internal action space, and storing the pattern as a reusable skill. The system also learns from the user's narration, understanding the intent behind each step. This multimodal approach—combining visual and verbal information—makes the learning more robust.

Comparison to other AI tools

Other AI assistants, such as Google's Gemini or Microsoft's Copilot, offer automation capabilities but typically require users to write or record macros in a scripting language. For example, Copilot can perform tasks in Excel or Word but requires the user to specify actions in natural language that then get translated into API calls. The demonstration-based approach of 'Teach Claude a skill' is more intuitive because it doesn't require any translation layer.

Similarly, robotic process automation (RPA) tools like UiPath and Automation Anywhere have long offered screen recording to create bots, but these are enterprise-grade platforms that require training and maintenance. Claude's implementation brings a similar concept to a consumer-level AI assistant, making it accessible to individual users and small teams.

By learning from a single screen recording, Claude can handle tasks that vary slightly each time, such as cleaning up a spreadsheet with different column layouts or organizing files with varying naming conventions. Traditional RPA often fails when the interface changes even slightly, but Claude's underlying LLM provides adaptability—it can understand the goal rather than just replaying a fixed script.

Potential use cases and applications

The range of tasks that can be taught to Claude is broad. Common workflows include:

• Invoice processing: Downloading invoices from email, extracting key fields, entering them into accounting software, and filing the PDFs.
• Image editing batch operations: Resizing, renaming, and moving images into folders based on content.
• Report generation: Pulling data from multiple sources, creating charts, formatting a presentation, and sending it to stakeholders.
• Data entry and cleanup: Copying data from web pages into spreadsheets, removing duplicates, and standardizing formats.
• Social media management: Scheduling posts across platforms, responding to comments with predefined templates.

For power users, this feature can dramatically reduce time spent on repetitive tasks. Instead of spending 20 minutes every day preparing a sales report, users can teach Claude once and let it do the work. The AI can run these skills on command or be triggered by events (e.g., when a new file appears in a folder), though that level of automation is not yet available.

Limitations and considerations

While the feature is promising, there are limitations. First, the skill learned is tied to the specific application and environment where it was recorded. If the screen layout changes or the target website updates its design, the skill may fail. Claude's adaptability can mitigate some of that, but not all. Second, the feature currently works only in the desktop app and requires the user to be signed in and running Claude Cowork. It does not run in the background unattended.

Privacy is another concern. Screen recording captures sensitive information, and although the data is processed by Anthropic's cloud servers, users must trust that it is handled securely. Anthropic states that recordings are encrypted in transit and at rest, and that the stored skills do not retain the original video—only the learned workflow pattern. However, users should avoid demonstrating tasks involving passwords or financial data until they are confident in the safeguards.

The feature is also limited to tasks that can be performed on screen. Physical actions, such as scanning a document with a hardware scanner, cannot be taught. Similarly, tasks that require interactions outside the desktop (like sending a text message through a phone) are out of scope.

Availability and rollout

Anthropic is rolling out the feature gradually to Claude Pro, Max, and Team subscribers through the desktop app's Cowork interface. The company plans to expand to more tiers and platforms based on feedback. Users on the Pro plan have reported the feature already appearing, with others expecting it in the coming weeks.

To access it, ensure the Claude desktop app is updated to the latest version. Then, as described, open Cowork mode and select 'Record a Skill'. After recording, the skill appears in a library where it can be renamed, edited, or deleted. Skills can be triggered by speaking or typing commands like 'Run my invoice skill'.

Anthropic has a history of iterative feature releases. Earlier this year, Claude gained the ability to use external tools via APIs, and now this demonstration-based learning. The trajectory suggests that future updates will allow skills to be chained together—for example, a skill that first downloads a report, then another that emails it—and eventually run autonomously on a schedule.

For anyone who performs the same workflow every day, this could be one of Claude's most practical upgrades yet. It transforms the AI from a passive responder to an active participant in daily tasks, learning not just from data but from human demonstration. The days of meticulously crafting prompts may be numbered.

Source:Android Authority News