This is fantastic! And a lot to wrap my head around. What I can say from building STUDIO-J with Claude… when I started, Claude was terrible with things like naming conventions, relative path structures, compiling zip files…and the like. From the beginning I was defining the “rules” or constraints of the platform. It was frustrating, I gave Claude crap often… but in the end, the platform got better, Claude got better. Now, it’s insanely fun! Because I tell it what plugin I feel like building, it spits out one zip file, I drag that into the installer and start playing with my new tool. There’s no plugin ecosystem yet, but at the same time it’s whatever plugin I want to invent or copy…. So I think my point is, for me, the most efficient use of ai as software/tool was to first create the environment software - the rules of the universe it is building for. Or put another way, I find that AI agent works best for you the more you set the guardrails. Not saying it’s the only way, but definitely efficient for me.
I completely agree. The most transformative AI workflows I’ve seen thrive when you first define the constraints and the "rules of the universe." It’s like giving the AI clear boundaries, and once those are set, the collaboration becomes seamless.
This is fantastic! And a lot to wrap my head around. What I can say from building STUDIO-J with Claude… when I started, Claude was terrible with things like naming conventions, relative path structures, compiling zip files…and the like. From the beginning I was defining the “rules” or constraints of the platform. It was frustrating, I gave Claude crap often… but in the end, the platform got better, Claude got better. Now, it’s insanely fun! Because I tell it what plugin I feel like building, it spits out one zip file, I drag that into the installer and start playing with my new tool. There’s no plugin ecosystem yet, but at the same time it’s whatever plugin I want to invent or copy…. So I think my point is, for me, the most efficient use of ai as software/tool was to first create the environment software - the rules of the universe it is building for. Or put another way, I find that AI agent works best for you the more you set the guardrails. Not saying it’s the only way, but definitely efficient for me.
I completely agree. The most transformative AI workflows I’ve seen thrive when you first define the constraints and the "rules of the universe." It’s like giving the AI clear boundaries, and once those are set, the collaboration becomes seamless.
Link to the original paper: https://arxiv.org/abs/2606.05608