Claude Computer Use: I Used It to Research This Post
Computer Use is live. Here’s what actually happened when I tried it.
As soon as I saw the announcement, I went to Claude app, clicked Dispatch, and typed:
“Can you go to X (I’m already logged in on my computer) and find 2-4 posts worth talking about on my Substack?”
I was curious if it’s going to work, but it actually worked!
Claude opened Chrome, went to X, scrolled my feed, looked at the engagement numbers, and came back with this:
> “Here’s what’s popping on your feed right now - the dominant story is Claude computer use. Top posts worth covering:
> 1. @cryptopunk7213 - 540K views, 2.3K likes. “Anthropic “bad word…” killed it (again). biggest AI product launch of the year so far.” The angle for you: you’re literally using this right now. A first-person workflow breakdown would land perfectly.
> 2. @mikefutia - Claude Code + Google Stitch 2.0. Practical stack update: code plus design generation in one loop. Strong “copy-paste framework” angle.
> 3. @dabit3 - Devin now has a fleet of agents. The “agent minions” framing is sticky. Good counterintuitive angle: what changes about how you work when you can delegate to a swarm?
> 4. @damianplayer - OpenAI offering PE firms 17.5% guaranteed returns. Business angle, not a workflow piece, but speaks to the structural shift in AI monetization.”
Honestly, I was kind of blown away.
Then it asked if I wanted to start drafting.
Here’s what I didn’t ask Claude to do: browse the internet in general. I asked it to look at my X feed, while I’m logged in, with my following list, and pull out what’s worth writing about. It read what I would have read. It handed me the synthesis instead of the raw scroll.
That’s what Claude computer use actually is.
What Claude computer use did in 90 seconds
Here’s what happened when I ran that prompt.
Claude opened Chrome, navigated to X, and read my feed. It looked at engagement numbers, identified the dominant story of the day, and surfaced four posts worth noting.
For each one, it wrote a one-sentence editorial angle calibrated to Prosper specifically, looked pretty nice. I am assuming it checked the memory for what I write about and that’s how it gave me the angle.
So then it stopped and asked what I wanted to do next.
The cool thing was that X’s algo already filtered for me, so Claude was working with my actual feed, not a generic one. So now it can become a autonomous research assistant that has access to whatever you’re already logged into.
It included actual numbers. 540K views, 2.3K likes. It is useful for knowing whether something is a niche conversation or a main-character moment.
It gave me angles and told me how each one could become something for my audience. It’s nice to look at these things because it makes you think about various posts differently.
I appreciate you reading this far. If you’re finding this useful, you can subscribe below to get Prosper posts like this in your inbox.
How Claude computer use changes your research loop
Before today, here’s how I thought about AI for content research:
I read things. I watch videos. I form an opinion. I start writing. Claude accelerates the writing.
That assumed I had to do the reading. I was the input. Claude was the accelerant.
Computer use can change where Claude enters the loop.
What you still have to own: the judgment. You decide which of the four posts actually becomes a story. You decide whether the angle Claude suggested is right. You bring the opinion that makes the piece worth reading.
Claude can handle the information retrieval. You handle the thinking.
I prefer that we all keep our thinking hats on!
How to use Claude computer use for research today
This is just one of the ideas that immediately popped in my mind - using “Computer use” feature as a research assistant. I am sure there are many other ways to use it, and I am sure people online will talk about it more, probably me including.
Again, the amazing thing about this is that you can send the prompt straight from your phone. If you have Claude on your phone and the Claude in Chrome extension installed, this works right now.
The simplest version of what I did:
Can you go to [site you’re already logged into] and find [specific thing you’d
normally spend 15 minutes looking for]?Start narrow and simple. Not “research my industry.” More like: “go to my LinkedIn notifications and tell me who engaged with my last three posts” or “open my email and find threads from this week that still need a reply.”
The bigger picture: 4 ways Claude now works remotely
Computer use is one piece of something larger Anthropic has been quietly building out.
Right now, Claude has four distinct ways to interact with you remotely: Dispatch (trigger a full Cowork session from your phone, now with computer use built in), Remote Control (Claude controls your desktop independently), Channels (Claude in messaging apps like Telegram), and web sessions (Claude running in the browser). Each one is a different access point to the same underlying capability.
The pattern Anthropic is following: every couple of weeks, another constraint on when and how you can use Claude disappears.
Getting started with Claude computer use
If you’re not set up with Claude in Chrome yet, I wrote a guest post about it here. It’s a 5-minute install.
Once you are: try the prompt I used. Try something simple. Pick one site you’re already logged into. Ask Claude to find something specific.
FAQ
Q: Who can use it and what do I need?
Claude Pro and Max subscribers only, for now. macOS only. You need to enable it in the desktop app settings, and your desktop app needs to be awake and running. Once that’s done, you can pair it with the mobile app via Dispatch and hand off tasks from your phone.
Q: How does Claude decide when to use computer control vs. a connector?
Connectors come first. If Claude has access to a Slack or Google Calendar connector, it’ll use that. Computer control kicks in when there’s no direct integration - that’s when Claude takes over the browser, mouse, and keyboard to complete the task itself.
Q: Does Claude just start clicking around on its own?
No. Claude always asks for explicit permission before accessing a new application. You can stop it at any point. According to Anthropic, it also scans for prompt injection automatically in the background.
Q: What are the current limitations?
It’s a research preview, so it won’t always work perfectly. Complex tasks sometimes need a second try, and working through your screen is slower than a direct integration. Anthropic recommends starting with apps you trust and avoiding sensitive data for now. Some apps are off-limits by default. macOS only for now - Windows support isn’t available yet.
This is awesome! I'm admittedly quite behind with non-code use cases for AI, and I always appreciate reading your newsletter to catch up
Your loop diagram landed for me. The "you were the input, Claude was the accelerant" framing is exactly how most people are still thinking about it. And it's already outdated as of today.
The pre-writing intake problem is a real doozy. I publish daily and the writing isn't what eats the clock. It's the 99 minutes of scrolling before I know what's worth writing about. Having Claude handle that while keeping editorial judgment on my side of the table is the split I've been trying to build with automation. (With mixed results. Mostly mixed. Fine, mostly bad. But educational.)
Curious: have you tried it on platforms beyond X (not a huge fan of that site)? Wondering if it handles less algorithmically sorted feeds as well.