Anthropic's 'Record a Skill' turns screen recordings into reusable Claude Cowork automations

Announced July 21, 'Record a Skill' targets the biggest friction in agentic AI: describing what you want. Instead of writing detailed prompts, users record themselves performing a workflow on screen; Claude analyzes the recording and generates a reusable automation that can be rerun on demand, all inside the shared Claude Cowork workspace. It is available to Pro, Max and Team tiers.

The design bet is that demonstration beats specification for repetitive knowledge work — a 'show, don't tell' interface for building agents. That resonates with a broader sentiment Andrej Karpathy voiced this week (29K likes) about wanting lower-effort ways to hand LLMs enough context, e.g. rambling by voice rather than typing. Record-a-Skill is the same instinct applied to workflows.

Competitively it slots against OpenAI's ChatGPT Work agents (Greg Brockman highlighted cloud-run agents that work 'with your laptop closed') and AWS's AgentCore push. The differentiator is capture: turning an observed demonstration directly into a durable, shareable skill within a team workspace.

Watch adoption and reliability — screen-recording-to-automation is easy to demo and hard to make robust across real, messy workflows. Whether recorded skills generalize beyond the exact recorded path is the open question.