Everyone’s talking about Claude right now, so I downloaded Anthropic’s new Claude Cowork research preview and basically let it control my Mac (yes, mildly terrifying). Think of it like Claude Code for normals: a real app with a more approachable interface that treats each chat like a task with trackable steps, can read/write local files, and (optionally) browse via the Chrome extension. It’s technically sandboxed without root access, but I still wouldn’t let any agent run fully autonomous on my main machine unless I had it in a virtual environment—or on a totally separate “AI robot” computer. That said, my Mac is backed up daily to my NAS, so… we yolo’d. I tested a few real workflows: auto-sorting photo folders into RAW vs JPEG (including multiple RAW formats across cameras), and then the spicy one—organizing my chaotic desktop while explicitly not touching a “stuff” folder. Cowork crushed the quantity game here: it made clean “keep” vs “can delete” buckets for screenshots and screen recordings, plus documents, audio, design files, and misc. Then I tried turning a wall-of-text doc into a slide deck; it produced a decent starter PowerPoint, but hit limitations like no direct Google Slides integration and some wonky formatting. My biggest gripe: Cowork (and Claude generally) can’t generate transcripts from audio/video—you have to bring transcripts yourself. Overall, it’s not life-changing today, but it’s a very real foundation for agentic workflows that normal people can actually use.