I also like watching through the thinking that happens when I have Claude executing asks. You start to get a sense for the flow and how it all comes together which makes it easier to ask for what you want next time.
Memory is also the piece that I haven't really solved. My solve so far is to just create a second brain database in Notion that Claude can connect to, and that I add to all the time (I've just saved this post to it). It seems to work pretty well.
Finally, a concept that inspires me enough to dig deeper into agents, and explains in such a way that my semi-technical mind doesn't immediately walk away from in mild terror. Thanks, Ben!
I have separate folders for areas or projects. So writing is a folder with sub folders for specific topics. They are grouped this way because I have general guidelines all writing shares. I used to be specific about file names and what’s in them, but not anymore. I just tell Claude or Codex what I want to do and let it manage the file system it needs. Like I want you to always remember these writing rules and use these as writing samples to guide you. It figures out how it needs to structure things. If I find a hole in it’s thinking I just tell it to fix it. I tell it I want it to remember and have history and it sets up git. I do it this way because I got tired of operating at files level and want it to take care of that. Maybe I give up something this way but it seems to work. I created agents in Openclaw to learn that usage. Each agent had its own instructions and model to use. It took a lot of effort to set up and get working correctly and to keep it on track. I don’t see the need for all that effort in Claude or Codex if I am only using one model. I may think about it again at some point. Even if I do it I think I will first tell Claude or Codex to set up agents that would be helpful to share across projects and let it decide. The tools are so smart I think it’s better use of my time to give them rules and guidelines and let them decide.
I also like watching through the thinking that happens when I have Claude executing asks. You start to get a sense for the flow and how it all comes together which makes it easier to ask for what you want next time.
Memory is also the piece that I haven't really solved. My solve so far is to just create a second brain database in Notion that Claude can connect to, and that I add to all the time (I've just saved this post to it). It seems to work pretty well.
Finally, a concept that inspires me enough to dig deeper into agents, and explains in such a way that my semi-technical mind doesn't immediately walk away from in mild terror. Thanks, Ben!
❤️❤️❤️
I have separate folders for areas or projects. So writing is a folder with sub folders for specific topics. They are grouped this way because I have general guidelines all writing shares. I used to be specific about file names and what’s in them, but not anymore. I just tell Claude or Codex what I want to do and let it manage the file system it needs. Like I want you to always remember these writing rules and use these as writing samples to guide you. It figures out how it needs to structure things. If I find a hole in it’s thinking I just tell it to fix it. I tell it I want it to remember and have history and it sets up git. I do it this way because I got tired of operating at files level and want it to take care of that. Maybe I give up something this way but it seems to work. I created agents in Openclaw to learn that usage. Each agent had its own instructions and model to use. It took a lot of effort to set up and get working correctly and to keep it on track. I don’t see the need for all that effort in Claude or Codex if I am only using one model. I may think about it again at some point. Even if I do it I think I will first tell Claude or Codex to set up agents that would be helpful to share across projects and let it decide. The tools are so smart I think it’s better use of my time to give them rules and guidelines and let them decide.