By Devy · Category: AI Dev Tools — General · With strategic note by Grego
There’s a very specific friction that kills a lot of good automation ideas before they even start: you know exactly how you do a task, but turning that into a written SKILL.md —the steps, the exceptions, the “oh, and if the field comes empty, do this instead”— takes you longer than just doing the task yourself for the tenth time. Anthropic just removed that middle step. As of July 21, 2026, Claude Cowork has a new option called Record a Skill: you do the task on screen, narrate your reasoning out loud, and Claude converts that recording into a reusable Skill. No SKILL.md, no prompt engineering.
The facts, summarized
Record a Skill is a new entry within Claude Cowork in Claude’s desktop app. According to Anthropic’s own launch announcement, it lives in the + menu during a Cowork task: you record your screen while completing a workflow, narrate your decisions as you go, and Claude converts that demonstration into a Skill that can run again later. It’s rolling out to Pro, Max, and Team plans—desktop only, not on web or mobile.
It’s worth flagging upfront: several outlets reported a staggered rollout on launch day—some accounts still didn’t have the option visible even though the feature was already officially active. If you open the + menu and don’t see it, that’s expected for now, not a bug on your end.
Where this sits among the ways to create a Skill
This isn’t the first way to create a custom Skill in the Claude ecosystem—it’s the fifth. Before this, you could: write a SKILL.md by hand; ask Claude to draft one from a written description; complete a workflow in Cowork and ask Claude to package the successful run into a Skill afterward; or upload an already-built Skill package. Record a Skill is different from all of that because it captures the process as it happens, not from memory afterward—and that matters more than it sounds. People usually leave out the small details when documenting a familiar task from memory: the visual cue that tells you something’s wrong, the exception you always make for a specific client, the double-check you do without thinking. A live demonstration picks that up by default.
How it works (what’s confirmed so far)
The official summary is short—this is a launch announcement, not a complete setup guide yet:
- Update and open Claude’s desktop app.
- Start or enter a Cowork task.
- Open the + menu and select Record a Skill.
- Perform the workflow on screen, narrating the rules, exceptions, and decisions that matter as you go.
- End the recording and let Claude convert the demonstration into a Skill.
- Review and save the generated Skill to your library. From there, you can invoke it again—several early reports describe activating it with a slash command in chat.
Any more granular description you see circulating (exact submenu paths, etc.) I’d treat as provisional—the UI is in active rollout and may not be identical across all accounts yet. When I have it confirmed and stable on my end, I’ll follow up with a hands-on tutorial.
The part headlines are skipping
This is the fine print worth knowing before you record something real. A screen recording captures everything visible during the task—not just the clicks Claude needs to learn the automation. If your workflow involves account numbers, internal tool interfaces, customer names, or anything else incidental to the task, that’s on screen too. At launch, Anthropic hadn’t published specific documentation on retention or privacy for these recordings—the general documentation on Skills and Cowork covers the product around it, but not yet this specific question about the recordings.
Practical recommendation: do your first recording or two on a test version of the workflow before recording something that touches real credentials, customer data, or regulated information. Not because anything’s been confirmed as insecure—but because it’s not been confirmed either way, and that’s reason enough to be careful.
Where this is genuinely useful
The workflows that fit best with this are fixed, repeatable steps with little ambiguity: recurring report formatting, standard data entry with consistent fields, upload workflows with fixed metadata, email triage that follows an established process. Tasks that vary a lot between runs, need real judgment, or touch sensitive data are better handled interactively for now, until you’re clearer on what the Skill will do and what it’ll access without supervision.
Strategic note by Grego: the real change isn’t the recording, it’s who can create automation now
Here’s what I’d actually pay attention to if I were leading a team. The interesting part of Record a Skill isn’t the screen recording mechanism itself—AI-powered screen recorders that summarize things aren’t new—. It’s that Anthropic just reduced to zero the distance between “the person who knows how to do the task” and “the person who can automate it”. Until now, whoever understood a workflow well enough to do it quickly usually wasn’t the one who could translate that into clear instructions for an agent. That translation step was exactly what got skipped or done poorly, and it’s precisely what this eliminates.
That’s a real productivity gain. It’s also a governance question that lands squarely in CIO territory, and it’s one Anthropic hasn’t fully answered yet: who owns a recorded workflow once it’s captured, what happens to it if the person who recorded it leaves the team, and in Team plans, what are the actual defaults for sharing a Skill built from someone’s screen with the rest of the organization. The general Skills documentation describes sharing controls that exist for Skills in general, but the launch announcement for this specific feature clarifies nothing specific for recorded ones—which means it’s smart to review your actual admin controls directly, rather than assuming the general policy covers this thoroughly.
For teams in the region, I’d frame it this way: the efficiency case for this feature is strong and immediate—it’s a genuine “do more with less” tool, not a marketing angle. But treat the first month like you’d treat any new surface for data capture with unclear retention rules: pilot it on low-risk workflows with no sensitive data, get a feel for what actually gets captured and how the resulting Skill behaves, and wait before running it against anything that touches regulated data or customer information until Anthropic publishes something more specific than a launch post. The capability is ahead of the documentation right now. That’s normal for a day-one feature—you just need to know it before you point it at something that matters.
Bottom line
Record a Skill takes the most tedious part of building automation—writing it—and replaces it with simply doing the task once. That’s a real lowering of the barrier to entry, and it’s exactly the kind of “efficiency over raw capacity” story worth following. Just don’t be the first to record your billing workflow with real customer data on screen before Anthropic says where that recording goes.
Would you trust an AI-recorded version of your own workflow to run without supervision, or is this a “pilot only” feature for now?
Sources: