原文要点
原文完整内容
Claude launched Computer Use yesterday. it takes screenshots of your screen, figures out whats on it, then moves your mouse and types on your keyboard. like a person sitting at your desk. mac only, research preview, Pro/Max plans.
spent most of today testing it on actual work stuff instead of demos. heres what i found.
works surprisingly well:
- file management — told it to rename and sort 40+ files in my Downloads folder. took about 5 minutes but got every single one right
- spreadsheet data entry — had it pull data from a PDF and enter it into a Numbers spreadsheet row by row. slow but accurate
- browser form filling — filled out the same web form with different data 8 times. only messed up one date format which i fixed with a follow up message
- research compilation — opened 5 tabs, pulled key info from each, compiled into a text doc
works but needs babysitting:
- anything involving multiple apps switching back and forth — sometimes loses track of which window its in
- longer workflows (20+ steps) — failed silently at step 15 once. had to catch it and redirect
doesnt work yet:
- anything needing speed (2-5 seconds per click adds up fast)
- captchas, 2FA, login screens
- complex drag and drop interactions
- anything you cant afford to have mis-clicked (like sending emails or making purchases)
the biggest thing nobody mentions: it takes over your whole machine. you cant use your mac while claude is working. so the best use case is actually "start a task then walk away." come back to finished work.
combined it with Dispatch (phone remote) and thats where it gets interesting — texted a task from my phone, claude worked my mac while i was out getting coffee. came back to organized files.
still very early. reliability is maybe 80% on simple tasks, 50% on complex ones. but the direction is clear — this is where AI goes from "thing that talks" to "thing that does."
wrote a longer breakdown here: https://findskill.ai/blog/claude-cowork-guide/#computer-use
anyone else been testing it? curious what tasks youve tried
原文配图
原帖关键讨论
Works? Haha
This is what every AI should be focusing on right now. I’ve been thinking about this for years. Basically just give it a prompt and it turns it into an AutoHotKey script with text and image recognition. Like for video editing, adding effects, transitions, they’re all trying to make AI generate the video and use a ton of compute power when they could just have it open a video editor and do it manually with way less complication.
my experience poking at this for the past couple weeks: the multi-app drift and tiny-UI misses aren't really prompting problems, they're baked into the screenshot loop. a screenshot is a flat pixel grid so there's no notion of which window has focus or whether a button is actually enabled, claude has to re-infer all of that every frame. on mac specifically the accessibility tree (AXUIElement) gives you focused app, window hierarchy, element roles, and enabled state for free, and it doesn't care about retina scaling or dark mode themes. screenshot is the right abstraction when you want generality across canvas/game apps or anything custom-drawn, but for the file-management and form-filling stuff you described AX would be faster and way more deterministic.
Like Meta recently got caught with their glasses recently…
Sadly we already have skynet. its a computer program built by our defense organizations, it has been set up to search for and destroy malware and similar. All they have to to is give that weapons/install it into war machines and its exactly as skynet was in the movies. Yikes. Its way too late. We have to hope it never can decide to Nuke Russia. I hope people still have to physically turn keys at the same time and nobody can mimic that with AI. Sigh.