Funny you mentioned GTD, I went through a similar reflection on a recent post about how AI agents increase how much we can do, but also require supervision/management. GTD came to mind and I mentioned it there. I’m testing something based on LLM wiki to try to organize everything (personal + work/research). So far it’s been working well, with Obsidian being useful to navigate/edit (on top of telling agent to do it) and can also do some visualization.
Good idea about obsidian — I also use it but I hadn’t thought about it being my code dashboard. That’s a great idea. Maybe that can build the visualization tool I need.
I saw that, too. I think I'm going to start experimenting with the html workflow as I build out this new broader system. THat's a good chance to do that because what I read is you can do a lot more with html, and you also can create shareable documents a lot easier.
Scott, your "lifecycle of context" framing (intent → spec → execution → verification → persistence) is close to the object I've been writing toward. I think of it as a research factory: a system around the model that pins three things: the spec defining what's being done, the evidence the spec is revised against, and the process governing how the whole thing evolves. In my research I apply this to economic measurement from text. Your line about tools being endogenous to the philosophy is what kept me from shipping more individual skills, and it's why the factory's first job is to fix the property-rights question you raise: the researcher owns the spec and its revisions; the model executes within the spec; the factory makes that split inspectable rather than ambient. I'm piloting the general approach with colleagues at the Philly Fed, and giving a talk at NBER's Innovative Information Initiative on May 29 that introduces the research factory idea and the elementary pieces: static specs and the dynamic beats that revise them.
Funny you mentioned GTD, I went through a similar reflection on a recent post about how AI agents increase how much we can do, but also require supervision/management. GTD came to mind and I mentioned it there. I’m testing something based on LLM wiki to try to organize everything (personal + work/research). So far it’s been working well, with Obsidian being useful to navigate/edit (on top of telling agent to do it) and can also do some visualization.
Good idea about obsidian — I also use it but I hadn’t thought about it being my code dashboard. That’s a great idea. Maybe that can build the visualization tool I need.
There is also a good case for HTML as the main medium for human-facing AI outputs https://x.com/trq212/status/2052809885763747935
I saw that, too. I think I'm going to start experimenting with the html workflow as I build out this new broader system. THat's a good chance to do that because what I read is you can do a lot more with html, and you also can create shareable documents a lot easier.
Scott, your "lifecycle of context" framing (intent → spec → execution → verification → persistence) is close to the object I've been writing toward. I think of it as a research factory: a system around the model that pins three things: the spec defining what's being done, the evidence the spec is revised against, and the process governing how the whole thing evolves. In my research I apply this to economic measurement from text. Your line about tools being endogenous to the philosophy is what kept me from shipping more individual skills, and it's why the factory's first job is to fix the property-rights question you raise: the researcher owns the spec and its revisions; the model executes within the spec; the factory makes that split inspectable rather than ambient. I'm piloting the general approach with colleagues at the Philly Fed, and giving a talk at NBER's Innovative Information Initiative on May 29 that introduces the research factory idea and the elementary pieces: static specs and the dynamic beats that revise them.
Brilliant. I am committed to doing the same.