Cowork reached my phone this week. I still can't tell where chat ends and Cowork begins.

This week the AI tools I lean on rearranged themselves under me. Anthropic pushed Claude Cowork onto web and mobile — it had been desktop-only since January — so I can now hand it a task from my phone and pick the result up at my desk. Days later OpenAI shipped its own answer, ChatGPT Work, an agent that does more or less the same job from inside the chat box you already use. I've spent the week rewiring how I operate around all of it, and I want to be honest about something: it's powerful, and it's genuinely confusing.

Two of the lines make complete sense

The split between Claude Code and Cowork I understand immediately. Code lives in the terminal, it's for people building software, and its output is software. Cowork lives in the app, it's for knowledge work, it runs in an isolated machine in the cloud, and it works across your files and your connected apps — you hand it an outcome and it produces the deck, the spreadsheet, the research. Same underlying engine, two different jobs. That distinction is clean, and I use both happily depending on whether what I want out the other end is code or a document. It's the same clarity I wanted when I wrote about the difference between skills and agents.

One of the lines doesn't

Chat versus Cowork is where I lose the thread. Chat can already reach into my connectors and pull things out of them. Cowork uses the same connectors, but it can go away and do multi-step work with what it finds. So the honest question I can't answer cleanly is: when do I open one instead of the other? The boundary is vague, and I don't think I'm the only one squinting at it — the coverage of the mobile launch noted that most Cowork users aren't doing anything that looks like coding at all. The audience is knowledge workers who've been told there are two doors and haven't been told which one to walk through.

The mobile launch exposed the seams

Putting Cowork on the phone is exactly the right move, and it's also where the whole thing stopped fitting together for me. The promise is continuity: start anywhere, finish anywhere. In practice, in these first days, it only ran one way. A session I kick off on my Mac doesn't appear on my phone. A session I start on my phone I can pick up in the browser and carry on — the same way I already move between devices in Claude's design tool. It's early, it's beta, and it'll settle. But right now the picture in the marketing and the picture on my two screens don't quite match.

The tools aren't the hard part anymore. Knowing which surface to point at which problem is.

Why I can exploit the mess anyway

Here's the part that matters more than any single feature. Because Cowork runs in the cloud, it can only be as good as the context it can reach — so I mirrored my key skills and workbenches to a connected cloud drive it can read. Now when I start something, the agent pulls the right skills and the right project context instead of beginning from nothing. The leverage isn't the app. It's the plumbing I built behind it, which is the same lesson I keep relearning about knowing which machine you're actually standing on and about teaching an assistant to prep my week so it stops starting over.

That's also why I'm not thrown by OpenAI collapsing all of this into one surface. ChatGPT Work keeps you in the chat you already know and simply lets it go do the work — connectors, files, even driving your computer. There's a real argument that one door is less confusing than Anthropic's two, and I'm testing whether that's actually better or just simpler. I like Anthropic's split between Code and Cowork enough that I'd keep it. It's the Chat-and-Cowork seam I want closed.

The reason I have time to sit inside this confusion and learn it properly is that the very same tools already make me efficient enough to afford the learning curve. That's the compounding move, and it's why a hands-on operator pulls ahead of someone who's only ever read about these products: the judgement about which tool to aim at which problem is the leverage, not the tool. My traffic tells me I'm not alone in wanting to work this out — the readers keep arriving on exactly these questions.

The lines will settle over the next few releases. I'd rather be fluent in the mess now than tidy and late.

Sources & further reading

External
Claude Cowork — Anthropic
ChatGPT is now a partner for your most ambitious work — OpenAI
Anthropic brings Claude Cowork to mobile and web — VentureBeat

Related posts
Claude Skills vs Claude Agents: the difference that actually matters
The hard part isn't Git. It's knowing which machine you're standing on.
I taught my assistant to prep my week. The hard part was making it stop starting over.
The Death of Generic AI: Why Deep Domain Expertise Is the Only Real Leverage Left