On July 21, Anthropic quietly switched on a new feature in the ”+” menu of Claude’s desktop Cowork app: Record a Skill. Record your screen, narrate what you’re doing, and when you stop, Claude distills the demonstration into a reusable Skill file.
This is not a small productivity tool. It may be the road that enterprise knowledge management failed to walk for twenty years.
From Writing Skills to Doing Skills
Until now, creating a Skill in Claude meant hand-writing a SKILL.md — Markdown laying out steps, trigger conditions, input and output parameters, metadata. Non-technical people had to think like process designers to pull it off. The barrier was translating tacit experience into explicit instructions.
Record a Skill skips the translation. It captures screen actions and voice narration simultaneously — and the narration is the key, because narrating naturally exposes judgment logic: which cases to skip, which format the client demands, which anomalous number needs human review. Those things almost never make it into an operations manual. A live screen recording keeps them all.
The hands-on flow is smooth: open the Claude desktop client → click the [+] at the bottom left of the chat box → Skills → Manage skills → Add → Record your screen. Grant screen and microphone permission, perform your full workflow once while narrating the implicit judgment calls, stop recording, and Claude analyzes and distills the Skill automatically. Save it and it’s callable via slash command. The feature is open to Pro, Max, and Team users.
Why Narrated Screencasts Are Worth More Than Saved Documentation
The deeper meaning runs three layers deep.
Layer one: individual efficiency. The old friction was abstract writing. The new friction is “do it once, say it once” — both daily behaviors for a practiced operator, designer, or analyst, requiring no new skills. Skill creation cost drops from “half an hour writing a doc” to “five minutes of screen recording.”
Layer two: organizational knowledge. Enterprises have burned money for decades on knowledge bases, SOPs, and best-practice libraries, and the pain never changed: experts won’t write docs, and nobody maintains the ones they do. Documents go stale the moment they’re written and jam the moment they need upkeep. A narrated screencast is an order of magnitude cheaper to produce, and the resulting Skill is both human-readable and directly executable by AI. More importantly, when the expert leaves, what remains is not an orphaned Word file but a running automation. The MakeUseOf case is telling: a freelance writer built a content-generation workflow in Claude Cowork during handover, and the client asked, “Since you’re leaving, could you set this up so we can run it ourselves?” A five-year client relationship was replaced by a folder of Skill files.
Layer three: the industry structure. This points at a bigger trend — AI is redefining what a “work skill” is. Harvard Business School research from March 2026 shows hiring down 17% for the most AI-exposed roles but up 22% for roles requiring human-machine collaboration. AI is not eliminating work; it is re-partitioning it. “Being able to do it” is no longer scarce. “Being able to teach AI to do it” is. Record a Skill pushes that shift one step further: your workflow itself becomes a programmable, distributable, executable asset. You don’t use a tool to do your work — your way of working becomes the tool.
The Frame: From Step Lists to a Triple
Traditional Skill creation captures process description — step 1, step 2, step 3. Record a Skill actually captures three kinds of information:
-
Operation sequence (What) — the clicks, inputs, and navigations on screen;
-
Judgment logic (Why) — the conditional branches and exception handling that surface naturally in narration;
-
Output standards (How Good) — the operator’s immediate quality feedback (“this layout is wrong,” “this data looks off”).
All three together make a genuinely reusable Skill. Miss the middle layer and the Skill is a rigid checklist; with it, the Skill knows when to flex. That is why screen-plus-voice beats screen-plus-text-annotation by an order of magnitude — voice naturally carries tone, pauses, and emphasis, which are exactly the most natural externalization of judgment.
What Happens Next
Short term (1–3 months): Skill counts explode — not because of marketing, but because creation cost dropped from “unnatural” (writing docs) to “natural” (do it, say it). Every desktop user will record their frequent workflows, which creates the next problem: discovery and quality control. When one user has dozens of Skills and a team has hundreds, how do you find the right one, or judge its quality? Anthropic will need a Skill marketplace or equivalent ecosystem infrastructure.
Mid term (3–12 months): Skill circulation becomes a business model. Skills are private today, but once they can be shared, traded, and rated, a “Skill developer” role appears — not a developer in the traditional sense, but “a person who can package workflows into AI.” For freelancers and consultants the deliverable changes: no longer a report or a proposal, but a Skill that runs and produces the report automatically.
Long term (12+ months): once Skill creation, distribution, and execution mature, enterprise AI deployment economics change structurally — from “buy an AI platform and let everyone figure it out” to “buy a Skill portfolio covering 80% of routine work.” That is the same evolutionary chain SaaS walked from development platform to app marketplace.
What You Can Do Now
-
If you have Claude Pro/Max/Team, try Record a Skill today. Pick a workflow you run at least three times a week and record it once. Feel the distance from “doing” to “Skill.”
-
Think about your judgment-logic moat. While recording, deliberately narrate every judgment — why you skip this step, why that data gets priority, when you would intervene manually. That is where your Skill’s real value lives.
-
Watch the sharing and reuse mechanisms. Skills are local today, but the ecosystem will catch up fast. Decide early: which team workflows deserve standardization? Which “one person’s craft” should become “team capability”?
-
Reverse the question: what in your job could never be taught to AI by any recording? Those tasks are probably your next career direction.
The real signal of Record a Skill: the capability gap between AIs is shrinking, while the gap between people who use AI well is widening fast. People who can record, narrate, and convert tacit judgment into executable process are acquiring a new lever.
The question is whether you intend to be one of them.
Cover image: Anthropic / Claude desktop
