I teach AI to business people through Claude Cowork. My Work OS is ~80% of what you describe:
1. Each project had a Workflows/ folder with WF-01, WF-02, etc => maps out biz processes for that domain, f.i Sales, Marketing, Clients
2. Skills are my autonomous agents
3. My memory is in kb.md topic files; either Knowledge/ inside the workspace or _Knowledge/ outside for global knowledge, f.i. Naming conventions global vs client data local
4. Skills generate files in Outputs/ => weekly of periodic sweep across workspace to decide if new knowledge should be added to the kb.
Many challenges I won’t mention here.
If interested, I’d love to connect and have a show&tell session. It feels like we could benefit from it. DM me!
This is a great setup, and the overlap is telling. Your WF-01/WF-02 workflow folders and the Knowledge/ vs _Knowledge/ split (local vs global) are the two decisions that make or break these systems. Naming conventions global, client data local, that's the right call.
The Outputs/ sweep is the part most people skip. Generating files is easy. The weekly pass that decides what actually earns a place in the kb is where the system compounds. That's the whole loop.
On the papers, honest answer: there's no single canonical one for the folder-native pattern. It's stitched from a few threads:
- Cognitive Architectures for Language Agents / CoALA (Sumers et al., 2023), the memory taxonomy: https://arxiv.org/abs/2309.02427
- Generative Agents (Park et al., 2023), the reflect-and-write-back loop your Outputs/ sweep is doing: https://arxiv.org/abs/2304.03442
The rest is Anthropic's writing on context engineering and Agent Skills, plus Karpathy's second-brain framing. The folder-native part is practice running ahead of the papers, which is why it's worth writing about.
On the show-and-tell: I'm buried across a few things right now, so I can't do a call, maybe in one month from now. But let's keep it here in the open, more people learn that way. What's the one place your system breaks most often? Curious if it's the same spot mine does.
Thanks for the detailed answer and link. It’s Sunday morning here in the San Francisco Bay Area and you gave me some good reading with my coffee :-)
Because it teach knowledge experts and consultants to build a Work OS on Claude Cowork only, my challenges are probably different than yours.
The biggest frustration by far: global instructions. Claude Cowork won’t update Cowork global instructions and use the cache. I probably spent 20+ hours on this, claiming victory and finally gave up. It’s a known Anthropic bug. Developers have it better because you laod the ~.claude/CLAUDE.md (different logic). Anyway, I am putting everything in the project http://CLAUDE.md with file references. I am a bit obsessed about what you put there -or what you don’t put actually.
My second biggest challenge: how do I make it dumb simple for my students to automate their day-to-day tasks and business process and help them build a Work OS that compound instead of drifting, and change their way of working without having to learn all the technical details behind AI and Claude. Easier said than done :-)))
Still working on this but in essence I am in the “click no code” camp for teaching.
I am betting big on skills as a “knowledge connector”. In essence, my skills can embed knowledge in references/ or retrieve global knowledge from a dedicated folder organize in topic (not yet a second brain, will get there). It’s the http://SKILL.md that helps Claude decide what to load.
It works great for students and because I ship builder and audit skills to for creating and managing skills (think skill-creator), it becomes a system.
Net-net: I try to build a compounding system (Work OS) people can build a business on, not a one-time hack.
Same coffee-on-Sunday energy over here, glad it landed.
I've been down your exact road. I've taught a lot of non-technical people to set this up, and did a fair amount of consulting around obsidian-second-brain: how to install Claude, how to actually use it when you've never touched a terminal in your life.
Here's where I landed, and it's a bit contrarian. I skip Cowork entirely and teach them Claude Code through the Claude desktop app instead. Day one is overwhelming, no way around it. The terminal, the files, the whole thing. But the moment it clicks, they're operating at a level Cowork can't reach, and they stop hitting the walls you're describing. The global-instructions cache bug is exactly that kind of wall, and it just doesn't exist on their side once they're on Claude Code.
The reason it works: it puts them on the same substrate you and I use. Plain files they own, real CLAUDE.md logic, skills that load predictably. So what they learn actually compounds instead of being trapped in a UI that fights them. The steep first week buys them a system they can grow for years.
Cowork lowers the first-day barrier but caps the ceiling too early. For people trying to build a business on this, I'd rather charge through the hard start and hand them the real tool.
Wow! This one is going straight to my favorites.
I teach AI to business people through Claude Cowork. My Work OS is ~80% of what you describe:
1. Each project had a Workflows/ folder with WF-01, WF-02, etc => maps out biz processes for that domain, f.i Sales, Marketing, Clients
2. Skills are my autonomous agents
3. My memory is in kb.md topic files; either Knowledge/ inside the workspace or _Knowledge/ outside for global knowledge, f.i. Naming conventions global vs client data local
4. Skills generate files in Outputs/ => weekly of periodic sweep across workspace to decide if new knowledge should be added to the kb.
Many challenges I won’t mention here.
If interested, I’d love to connect and have a show&tell session. It feels like we could benefit from it. DM me!
PS: can you share the official research papers?
This is a great setup, and the overlap is telling. Your WF-01/WF-02 workflow folders and the Knowledge/ vs _Knowledge/ split (local vs global) are the two decisions that make or break these systems. Naming conventions global, client data local, that's the right call.
The Outputs/ sweep is the part most people skip. Generating files is easy. The weekly pass that decides what actually earns a place in the kb is where the system compounds. That's the whole loop.
On the papers, honest answer: there's no single canonical one for the folder-native pattern. It's stitched from a few threads:
- MemGPT: Towards LLMs as Operating Systems (Packer et al., 2023), the memory-tier idea: https://arxiv.org/abs/2310.08560
- Cognitive Architectures for Language Agents / CoALA (Sumers et al., 2023), the memory taxonomy: https://arxiv.org/abs/2309.02427
- Generative Agents (Park et al., 2023), the reflect-and-write-back loop your Outputs/ sweep is doing: https://arxiv.org/abs/2304.03442
The rest is Anthropic's writing on context engineering and Agent Skills, plus Karpathy's second-brain framing. The folder-native part is practice running ahead of the papers, which is why it's worth writing about.
On the show-and-tell: I'm buried across a few things right now, so I can't do a call, maybe in one month from now. But let's keep it here in the open, more people learn that way. What's the one place your system breaks most often? Curious if it's the same spot mine does.
Thanks for the detailed answer and link. It’s Sunday morning here in the San Francisco Bay Area and you gave me some good reading with my coffee :-)
Because it teach knowledge experts and consultants to build a Work OS on Claude Cowork only, my challenges are probably different than yours.
The biggest frustration by far: global instructions. Claude Cowork won’t update Cowork global instructions and use the cache. I probably spent 20+ hours on this, claiming victory and finally gave up. It’s a known Anthropic bug. Developers have it better because you laod the ~.claude/CLAUDE.md (different logic). Anyway, I am putting everything in the project http://CLAUDE.md with file references. I am a bit obsessed about what you put there -or what you don’t put actually.
My second biggest challenge: how do I make it dumb simple for my students to automate their day-to-day tasks and business process and help them build a Work OS that compound instead of drifting, and change their way of working without having to learn all the technical details behind AI and Claude. Easier said than done :-)))
Still working on this but in essence I am in the “click no code” camp for teaching.
I am betting big on skills as a “knowledge connector”. In essence, my skills can embed knowledge in references/ or retrieve global knowledge from a dedicated folder organize in topic (not yet a second brain, will get there). It’s the http://SKILL.md that helps Claude decide what to load.
It works great for students and because I ship builder and audit skills to for creating and managing skills (think skill-creator), it becomes a system.
Net-net: I try to build a compounding system (Work OS) people can build a business on, not a one-time hack.
Same coffee-on-Sunday energy over here, glad it landed.
I've been down your exact road. I've taught a lot of non-technical people to set this up, and did a fair amount of consulting around obsidian-second-brain: how to install Claude, how to actually use it when you've never touched a terminal in your life.
Here's where I landed, and it's a bit contrarian. I skip Cowork entirely and teach them Claude Code through the Claude desktop app instead. Day one is overwhelming, no way around it. The terminal, the files, the whole thing. But the moment it clicks, they're operating at a level Cowork can't reach, and they stop hitting the walls you're describing. The global-instructions cache bug is exactly that kind of wall, and it just doesn't exist on their side once they're on Claude Code.
The reason it works: it puts them on the same substrate you and I use. Plain files they own, real CLAUDE.md logic, skills that load predictably. So what they learn actually compounds instead of being trapped in a UI that fights them. The steep first week buys them a system they can grow for years.
Cowork lowers the first-day barrier but caps the ceiling too early. For people trying to build a business on this, I'd rather charge through the hard start and hand them the real tool.