Claude Cowork Now Learns Skills From One Screen Recording

Anthropic just replaced the hardest part of automating a workflow — explaining it to an AI in words — with something almost anyone can do without writing a single prompt: show it.

What Happened

The Decoder reports that Anthropic shipped a “Record a Skill” feature inside Claude Cowork on July 21, 2026. Users record their screen while completing a task and narrate what they’re doing out loud. Claude Cowork captures the screen activity, clicks, typing, and voice together, then converts the whole sequence into a reusable, rerunnable skill — no manual documentation or prompt engineering required.

The feature lives behind the ”+” menu in Cowork, the workspace section of the Claude desktop app, and is available now to Pro, Max, and Team plan subscribers. Once captured, a skill lands in the user’s skill library, where it can be triggered again the next time the same task comes up, rather than requiring someone to re-walk Claude through every step. The Decoder notes the approach mirrors a similar workflow-recording capability already shipped in OpenAI’s Codex, putting Anthropic and OpenAI on parallel tracks toward the same idea: teaching an agent by demonstration rather than description.

Why It Matters

The mechanism matters more than the convenience. Every prior approach to giving an AI agent a repeatable capability — custom instructions, system prompts, documented SOPs — required someone to translate a physical workflow into language first, then hope the model interpreted that language correctly. Record a Skill removes the translation step entirely. The demonstration is the specification. Claude infers the structure, the branching logic, and the intent directly from watching the task get done and hearing it explained in real time.

That has a direct effect on who can build automation. Prompt engineering has functioned as a soft skill barrier: people who are fluent in describing processes abstractly get more out of agentic tools than people who simply know how to do the process. A screen recording and a voice-over collapses that gap. The operations coordinator who reconciles the same spreadsheet every Friday doesn’t need to articulate the logic as a prompt — they just need to do the task once, out loud, and Claude does the abstraction work.

It also changes where automation gets captured. Institutional knowledge about “how we actually do this” typically lives in one person’s head, or in a stale wiki page nobody updates. A skill recorded directly from the task, at the moment it’s performed, is closer to ground truth than any documentation written after the fact — and it stays a living, executable artifact rather than a static writeup.

What This Means for Builders

For teams already running Claude Cowork, this is a strong argument to build a habit of recording high-frequency, low-judgment tasks the first time someone new learns them, rather than writing a how-to doc that will drift out of date. It’s a natural complement to the coordination and administrative work that already dominates Cowork usage — the status report consolidation, onboarding checklists, and spreadsheet reconciliation that make up roughly a third of all sessions are exactly the kind of repetitive, well-bounded work that benefits most from a record-once, run-forever skill.

For platform and enablement teams, the open question is governance: who reviews a skill before it enters a shared library, how permissions on a recorded skill get scoped, and what happens when the underlying tool or interface the skill was recorded against changes shape. Anthropic hasn’t detailed that layer yet, and it’s the part enterprise buyers will ask about first.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.