On July 21, 2026 (local time), Anthropic added "Record a skill" to Claude Cowork. Users record their screen while working and verbally explain the intent behind their actions, and Claude converts this into a Skill that can be reused for the same task. This update removes the file structure and syntax requirements that were previously needed to create custom Skills, shifting the means of teaching tasks from written procedures to live demonstration. However, task automation isn't complete the moment recording ends. It only becomes a team asset once you decide how to verify the generated procedure and who approves it for sharing.
"Record a skill" for Pro, Max, and Team
According to Claude's official X account, "Record a skill" is launched from the "+" menu in Claude Desktop's Cowork. Users record their screen while performing the target task themselves, verbally explaining what they're doing. Claude converts the recording into a Skill that can be executed again later. The feature is available for the Pro, Max, and Team plans; Free and Enterprise are not included in this announcement.
That's the extent of what's confirmed so far. Anthropic hasn't explained whether the recording is saved as a macro that replays click coordinates in sequence, or whether the meaning of the actions is extracted and reconstructed into a procedure. The structure of the generated files, and whether users can edit or export the content, also haven't been disclosed. It would be premature to interpret the claim that "the same task can be re-executed" as a guarantee of identical results even when the screen state changes.
Even so, this represents a significant change in input method. Previously, tasks had to first be put into words and then converted into instructions Claude could read. With the new feature, the actual screen and explanation can be provided as material simultaneously. This reduces the burden of having to recall, from a blank procedure document, the sequence and checkpoints that workers normally follow without conscious thought.
Demonstration before skill.md
In Anthropic's custom Skill creation guide, a Skill is defined as, at minimum, a directory containing skill.md. The YAML frontmatter at the top of the file must include a name and description. Reference materials or executable scripts can be added as needed, the folder is packaged as a ZIP, and uploaded via Claude's "Customize > Skills." The description text also serves as a cue for when Claude decides to invoke that Skill.
This approach makes content easy to review and allows code and reference materials to be incorporated. On the other hand, for non-technical staff transferring their daily work, the process of writing down implicit procedures and packaging them in the correct folder structure comes first. "Record a skill" instead receives this initial material from everyday work rather than from a completed document. It's a design that expands who can create Skills, from those who can write configuration files to those who know the actual work.
For example, in a task of transcribing invoices into a spreadsheet, what determines the outcome isn't the order of cell entries but which items are treated as authoritative and where exceptions should halt the process. Screen recording alone shows the operations but tends to lose the reasoning behind decisions. Adding voice explanation lets the worker convey even "why they checked that particular field." Whether Claude can properly separate actions from judgments, and extract fixed rules from inputs that vary each time, will determine the quality of Skills derived from recordings.
How Far Can Recording Reproduce the Task?
For existing custom Skills, Anthropic recommends checking the description, reference files, and dependencies before uploading, and testing how the Skill is invoked with multiple inputs after adding it. Even with recording as the new entry point, this verification process remains necessary. A demonstration can show one successful example, but that doesn't necessarily mean exceptions—such as blank fields, insufficient permissions, or changes to the screen layout—have been automatically defined.
In initial testing, you'd confirm whether the Skill is invoked for the intended request. Next, you'd swap in values that change each time, such as dates or customer names. Finally, you'd check whether it can safely stop when there's insufficient input or an operation fails. Going through these three stages early can reveal whether the Skill is merely a literal copy of the recording or is missing judgment criteria.
Being able to read the generated output is also essential for verification. Skills in general can include external packages and scripts in addition to instruction files. Anthropic itself asks that, before enabling a Skill with low trust—including one shared by a colleague—users check its content and dependencies. How granularly can the Skill generated by "Record a skill" be inspected and corrected? This is a specification worth confirming in the product interface before looking at reproduction-rate figures.
For Team Adoption, Decide on a Verification Path Before Sharing
Existing Skills for Team and Enterprise have a mechanism for sharing with specific colleagues or the entire organization. Sharing is disabled by default and must be enabled by an organization owner. Recipients receive a read-only copy, and when the creator updates it, the update is reflected for those it's shared with. However, whether Skills created from recordings can use this same sharing pathway hasn't been indicated in this announcement.
As the effort required to create Skills decreases, similar Skills are likely to proliferate across departments. On the operations side, you'll want a process where someone other than the creator reviews the content, and only tested versions get shared. You'll also need someone responsible for retesting after updates. A mechanism that automatically pushes updates to recipients makes distribution easier, but it also means erroneous changes spread just as quickly.
Drawing lines around which tasks to record also matters. Cowork can read and write connected files and operate the web and apps. In current remote sessions, work proceeds in an isolated environment on Anthropic's servers, and even when opening connected local files, processing happens server-side. Anthropic recommends setting up a dedicated working folder, restricting access to sensitive information, and having a human oversee the results. Operational practices that keep customer information or credentials out of screen recordings should be considered an extension of this same guidance.
Gaps also remain around audit functionality. According to Anthropic, Cowork activity is not currently logged to the Compliance API. Team and Enterprise owners can send events to a monitoring platform via OpenTelemetry, but it's unclear how much of the recording content or generation process for "Record a skill" falls under what gets logged. Whether the generated Skill can be viewed and modified, whether the storage and deletion of recording data can be managed, and whether the organization can track execution history—once these three points are addressed, it becomes possible to advance procedures created from demonstrations from a personal time-saver into a manageable organizational asset.