Your next skill starts with “watch me”

Some workflows are simple to perform and surprisingly awkward to explain. The clicks feel obvious while we are doing them; written instructions make every small decision visible.

Claude’s answer is Record a skill. Claude says we can perform a desktop task, narrate it as we go and turn that demonstration into a skill it can run again. The convenience is clear. The conditions for trusting the result are not.

The appeal begins with a familiar mismatch: we often know how to do a task without knowing how to turn it into a precise operating manual. Showing the work lets the sequence emerge from the work itself.

That changes where the effort sits. We may spend less time describing each step, but we still need to decide what Claude can see, what the demonstration reveals and how we will check the skill before trusting it again.

Claude announced Record a skill for Cowork in the Claude desktop app. It appears under the + menu and, according to Claude, is available on Pro, Max and Team plans.

That is the full confirmed perimeter. The post gives no pricing, geography or rollout detail, and it does not explain how a generated skill can be inspected, edited, constrained or validated.

Claude’s product image says the recording includes our screen, clicks, typing and voice before being sent to Claude and turned into a repeatable skill. That disclosure is central to the feature, not a footnote to it.

The same image shows “Saved as /file-expenses.” It is an illustrative result, not evidence that Creator Broadcast has tested the feature or that every recorded workflow will produce a dependable skill.

Your screen, clicks, typing, and voice are recorded, then sent to Claude and turned into a repeatable skill.

Claude-published product image

The announcement establishes a clear product proposition: demonstrate a task once and preserve it as something Claude can run again. It also identifies the desktop entry point and the three named plans.

It does not provide independent validation, a technical specification or evidence about performance. We can treat the post as proof of what Claude announced, but not as proof that the resulting skill will handle variation, recover from mistakes or reproduce the task consistently.

We perform the task on screen and talk through the choices. Claude says the screen recording and narration are sent to Claude, which turns the demonstration into a skill it can run again.

Claude does not describe verification as part of the feature. That final check is our safeguard: repeat the skill on harmless, controlled material and compare the result with what we intended to teach.

If Record a skill is available on our named plan, the first demonstration should use fictional or disposable material. A narrow filing or formatting task gives us an observable result without exposing sensitive work or creating costly consequences.

We should then run the skill again against a second controlled example and compare the output. This does not resolve the unanswered privacy or administrative questions, but it separates the appeal of the demonstration from the reliability of the saved behaviour.

The important shift is not that Claude can watch us work. It is that the demonstration itself can become the reusable instruction.

Creator Broadcast synthesis, grounded in Claude’s announcement

That could make skill creation more accessible to the people who know a process intimately but would never write a formal procedure for it. Their practical judgement can enter the system through the act of doing the work.

The same directness raises the stakes. A demonstration may contain more than the intended procedure, and a reusable result can preserve mistakes as easily as good judgement. Convenience becomes meaningful only when capture boundaries and verification keep pace.