Realizing very quickly the limitations of Claude Cowork...

I've been building the architecture for an AI Agency with the premise of utilizing Claude Cowork on and off for the past couple weeks now. Finally finished my "alpha" v1.0 and immediately started running into problems that I didn't think would be prominent until later on as the work got more complex.

First off, building parent-child relationships between Cowork projects is a bad idea lol. I have an Agency Central project as the main parent with individual Client projects as linked children. However, since instructions are native to the UI and not inherited naturally through the file architecture, this makes instructions communication difficult to say the least. One option is to only use the native UI project instructions for super minimal basic instructions across all projects, and then still write independent agency-project-instructions.md and client-project-instructions.md within the file architecture. However, this now requires a sync protocol to be established so that the agency project parent instructions are copied across all client project roots (in an About Me folder) which seems like an entirely unnecessary waste of tokens. Not to mention there's like 3 more .md files I've created (agency-profile, anti-ai-writing-style, and tool-router) which also need to be synced across all client projects as well. With Claude Code, instruction inheritance is already incorporated.

The other issue I have is with control within Cowork. You can't switch models, adaptive thinking, or effort levels mid-session. It requires a whole new session. Also, any connectors, plugin, or skills are ALL loaded up in every single session no matter whether you use them or not, which absolutely destroys token usage. Not to mention that I haven't even installed any yet and I'm still maxing out my 5hr usage limits on the Pro plan anyways. Seems like these are things Anthropic could easily change in Cowork, but to be honest idk shit about developing so I'm probably talking out of my ass. And if it's natively available in Claude Code, it would seem like I should probably be using that instead anyways.

So anyways, I've got Cowork re-writing its own architecture now to be implemented into Code. I still intend to use Cowork for scheduled tasks and maybe other specific tasks that are better suited for it (for example something that would require desktop use possibly, at least according to what Claude told me, but perhaps there's a way to do this in Code too) but overall 99% of the work I'll be doing from here on out will just be within Code.

Has anyone else run into similar problems with Cowork vs Code?

Thanks for coming to my TEDTalk.

1 comment