The sandbox conversation really clicks with me as someone who builds small web apps as side projects. The biggest friction for me isn't usually the coding itself, it's the environment setup and deployment. If I can spin something up, test it safely, and roll back when I inevitably break it, that removes a lot of the fear of experimentation.
Your point about open source contributions potentially increasing because of LLMs is something I've been thinking about. I maintain a small project, and honestly the barrier to contributing has always been more about environment setup than actual coding. If tools like this lower that barrier, that seems net positive.
I use Cursor for vibecoding projects, longform writing and now as a general project management suite. I keep meaning to kick the tires of Claude Code. I use Anthropic models through Cursor anyway. For writing, I use one repo per major fiction project. Set up CHARACTERS.MD, SETTINGS.MD, VIGNETTES. MD, etc. Aware of gaps I should consider?
Thanks for your review of Claude CoWork Simon. I’m a non dev wannabe founder, and to me, it seems quite a remarkable tool. I must admit I do worry about the prompt injection angle, and whilst Anthropic says ‘only let it use websites you know and trust’, I find this to be entirely impractical. It means you’d had to specify the urls of all sites like to be used, and at the beginning of a project, I just don’t have a sense of all the sites it might need to look at. I’m wondering therefore if the best form of sandbox might be buying another computer box engirley (a Mac mini perhaps), loading relevant files on that, and only using that for CoWork type work. The copying files across is kinda inconvenient, but at least if I get a prompt injection attack, it’ll only affect that subset of files and wreck what’s on that box, not my main system. I’m especially paranoid as I’m fully aware I have no idea what I’m doing, I just want to build prototypes fast for mock ups, to validate demand for apps in my industry vertical. If anything gets traction, I’ll have to scrap it and get someone who actually knows that they’re doing to build it properly, but the prototyping is so much faster now. I’m not testing can X be built, I’m testing should X be built - who will prove demand by giving me some money.
Great breakdown. What I love most about Cowork is that it shows real AI‑assisted workflows instead of toy demos. Seeing it handle tasks end‑to‑end feels like a glimpse of what “work with AI” will actually look like in practice
Cool! I somehow missed this. I’ve been trying to contort Claude Code with custom MCPs and just dumping stuff into folders. Hopefully this should make things easier.
Any sense as to what it would take (i.e., features) for you, as a developer, to change from using Claude Code to using Cowork? Can you see that ever happening or expect it to for a certain style of task?
Great overview of Cowork, thanks!
Feels like Code but as you point out as well, easier to use for most people.
(almost all non-tech people I've showed Code feel like the terminal is scary)
Great first look, Simon. The Cowork announcement validates the architecture pattern many of us have been exploring.
Anthropic built Cowork using Claude Code in just 10 days. That's the meta-story: AI agents building AI agents is now production-ready.
Anthropic's autonomous business experiment: https://thoughts.jock.pl/p/anthropic-project-vend-phase-2-ai-autonomous-business-lessons
The sandbox conversation really clicks with me as someone who builds small web apps as side projects. The biggest friction for me isn't usually the coding itself, it's the environment setup and deployment. If I can spin something up, test it safely, and roll back when I inevitably break it, that removes a lot of the fear of experimentation.
Your point about open source contributions potentially increasing because of LLMs is something I've been thinking about. I maintain a small project, and honestly the barrier to contributing has always been more about environment setup than actual coding. If tools like this lower that barrier, that seems net positive.
I use Cursor for vibecoding projects, longform writing and now as a general project management suite. I keep meaning to kick the tires of Claude Code. I use Anthropic models through Cursor anyway. For writing, I use one repo per major fiction project. Set up CHARACTERS.MD, SETTINGS.MD, VIGNETTES. MD, etc. Aware of gaps I should consider?
Thanks for your review of Claude CoWork Simon. I’m a non dev wannabe founder, and to me, it seems quite a remarkable tool. I must admit I do worry about the prompt injection angle, and whilst Anthropic says ‘only let it use websites you know and trust’, I find this to be entirely impractical. It means you’d had to specify the urls of all sites like to be used, and at the beginning of a project, I just don’t have a sense of all the sites it might need to look at. I’m wondering therefore if the best form of sandbox might be buying another computer box engirley (a Mac mini perhaps), loading relevant files on that, and only using that for CoWork type work. The copying files across is kinda inconvenient, but at least if I get a prompt injection attack, it’ll only affect that subset of files and wreck what’s on that box, not my main system. I’m especially paranoid as I’m fully aware I have no idea what I’m doing, I just want to build prototypes fast for mock ups, to validate demand for apps in my industry vertical. If anything gets traction, I’ll have to scrap it and get someone who actually knows that they’re doing to build it properly, but the prototyping is so much faster now. I’m not testing can X be built, I’m testing should X be built - who will prove demand by giving me some money.
Great breakdown. What I love most about Cowork is that it shows real AI‑assisted workflows instead of toy demos. Seeing it handle tasks end‑to‑end feels like a glimpse of what “work with AI” will actually look like in practice
Cool! I somehow missed this. I’ve been trying to contort Claude Code with custom MCPs and just dumping stuff into folders. Hopefully this should make things easier.
Any sense as to what it would take (i.e., features) for you, as a developer, to change from using Claude Code to using Cowork? Can you see that ever happening or expect it to for a certain style of task?