Prompt engineering just got a rival, and it involves talking to yourself while your cursor does the rounds. Anthropic has switched on a feature called Record a Skill inside Claude Cowork, and the idea is disarmingly simple: instead of describing what you want in text, you show it.
Here’s how it plays out. You hit record, run through a task on your screen exactly as you normally would, and narrate the steps out loud as you go. Claude watches the whole performance — the clicks, the sequence, the little decisions you rattle off — then digests it into a reusable skill. Next time the same chore rolls around, you hand it off rather than doing it yourself.
It’s a neat inversion of the usual AI relationship. Rather than you learning how to coax the model with the perfect wording, the model learns your way of doing things. For anyone who spends their week wrestling the same spreadsheet reformatting, report assembly or file-shuffling routine, that’s a genuinely different pitch. The repetitive stuff finally has somewhere else to go.
A few practical points are worth flagging:
- Where it lives: the feature sits in the Cowork interface of the Claude desktop app.
- Who gets it: Anthropic is rolling it out to Claude Pro, Max and Team subscribers.
- How it learns: screen recording plus your spoken narration, saved as a skill you can trigger again later.
The screen-plus-voice approach is the clever bit. Screen capture alone tells the model what you did; the narration fills in the why — the intent behind a step that might otherwise look ambiguous. That combination should, in theory, produce skills that survive small changes rather than breaking the moment a button moves.
Anthropic announced the feature on July 21, 2026, and it’s rolling out now rather than sitting in a distant roadmap. Pricing beyond the existing Pro, Max and Team tiers wasn’t disclosed at launch, so it’s unclear whether Record a Skill carries any usage limits of its own.
There are obvious open questions. How reliably does a recorded skill hold up when your app layout shifts, or when the data looks a little different from the demo run? How much narration is enough, and how much is too much? Teaching by demonstration is only as good as the demonstration, and messy real-world workflows are rarely as tidy as the tutorial version.
Still, the direction is telling. If Cowork can reliably turn a one-off screen recording into a dependable automation, it nudges Claude from a chat window toward something closer to a coworker who actually watched you do the job — and then quietly took it over.
