What this is
A Juejin developer logged a counter-intuitive experience this week: his MacBook sat closed at the office, yet Claude on his phone was busy blurring faces in a dance video. He had previously written a custom Skill for this task (think of it as "a set of steps you teach the AI"), calling on local tools like Python, ffmpeg, and OpenCV.
The mechanism is straightforward: Claude Pro and above users package their Skill as a zip and upload it to a personal container bound to their account, with paths like /mnt/skills/user/video-face-mosaic/. At execution, the Skill runs in a temporary Linux container in the cloud; the output (the video) is sent back to you, but your local photo gallery stays untouched—downloading to your phone is still on you.
Industry view
Supporters see this as the marker of an AI assistant that's truly "usable"—no longer just an office productivity tool, but one that is starting to enter daily life. The developer wrote: "From now on, whatever I need, I just walk through it once, save it, and I don't need a pile of apps." That matches the trend we've been tracking: AI is shifting from "conversation tool" to "execution tool."
But the cost ledger doesn't add up cleanly. Every time a Skill runs, Anthropic spins up a fresh Linux container to run Python and ffmpeg—the original author himself sighed that "A-co is still burning compute like crazy to expand market," and flagged that DeepSeek is about to raise prices while Anthropic is still spending cash to buy share. Then there's the privacy boundary: the cloud can process your video but can't reach your local photo library. This "cloud-execute + local-store" line is both a safety rail Anthropic set on purpose and a sign that, in the near term, the product can only cover lightweight, interruptible tasks—scalable monetization is still an open question.
Impact on regular people
For enterprise IT: AI assistant deployment cost shifts from "outfit every employee with a high-performance laptop" to "cloud compute + account subscription." Procurement will need to rebuild the math.
For individual professionals: white-collar workers can wrap repetitive workflows—weekly reports, statements, asset processing—into their own Skills and invoke them from a phone. One workflow no longer dies with a single device.
For the consumer market: the real rival to short-video, photo-editing, and clipping tools may no longer be another app, but the AI workflows users train themselves. The "build a tool" business logic is being quietly rewritten.