Awesome to see you back at it Brian! Go full agent workflow. No question about it. Claude Code or Codex are best right now. Gemini CLI is good but not as good. I setup a sandboxed Linux mini PC and installed Claude Code and OpenClaw with some throwaway accounts and it BLOWS my mind how good it is. Connected it to a burner WhatsApp account I can chat with to automate pretty much anything, including coding via chat message, rebuilding, commit/push to GitHub, etc etc. So much fun.
Crazy stuff. Yeah was talking to folks last night at a Xoogler meetup and OpenClaw was definitely one of the topics of the night. Definitely going to look at Claude Code. Seems to be where it's at.
Loved reading this. I feel like I either picked exactly the right time, or exactly the wrong time, to quit my software engineering job 18 months ago! I wonder if there's ever been such a rapid change in how computers are programmed before? My initial dabblings have followed a similar trajectory to yours: type a prompt into a generic chatbot and be amazed at the initial completely functional app that gets spit out; followed by uncertainty about the best way to refine / scale up from there. As you say, there are so many tools out there and new ones coming out each day, it's hard to know where to go... Hoping that reading this blog will give me the answer someday!
Thanks for checking it out. Yeah fun stuff. Definitely check out Johnkew Pickleball on YouTube if you want to go deep down the data driven pickleball paddle reviews. Fun stuff.
Or as I put it: minutes 1-10: "Whee, I'm flying!" hours 1-5: "...into a tar pit of despair."
Figuring out how to work around tooling limitations is definitely one of the most frustrating parts of this for me, especially since it often seems that if you just wait another few months, the friction has been largely overcome for many of the tooling problems you spent days trying to develop hacks to get around.
But that magic "Wow!" for the first 10 minutes keeps luring me back, and I'm slowly improving my workflow to the point of it actually feeling like a productivity enhancement most of the time now.
That is good to hear from you Jessica. If you think it's net positive then it must be net positive as you are not one to sugar coat. I totally hear you on the wait a few months, and I'm kind of glad I sat out for 9 months in that sense. Things are changing so fast where you read many suggestions of adding this or that to your prompt then later reading that those workarounds are all unnecessary with the latest greatest models or tools. The workflows definitely feel like they continue to rapidly evolve.
I can't speak to the external toolchains at all, and I haven't been trying to do any UI work. For me, the big turning point was when I stopped trying to use it in an IDE and really dove in on the CLI front. Going the conversational route to see what the agent thought it should do tended to help me figure out mistakes in both its planning and my own thoughts on what needed to get done much more quickly. Encouraging it to suggest things and ask questions before we really got going on the coding front helped me better think through what I thought I wanted, and also seemed to help me frame things in ways that the agent seemed to grok.
I've had plenty of mini "Ah ha!" moments that have helped me start using the tooling better, but I suspect the majority of them will age as well as milk. Maybe after I return from Japan I'll have to wipe my brain of all my accelerating hacks, clear my own context window, and begin again!
On the whole, as my eyesight has been continuing to glissade down with no self-arresting, I was settling into a certainty that coding was just not going to be possible for me (obviously there are plenty of low-vision and blind programmers out there, but at some point the tooling hurdle feels so large, that moving into architecture design and management feels far more tractable). The ability to leverage agentic support for coding has reawakened my sense of confidence in being able to do actual code development as a via career path again. I'm actually submitting code that contains more than just markdown, GCL, or BCL these days, squeezed in between the TLMing!
Awesome to see you back at it Brian! Go full agent workflow. No question about it. Claude Code or Codex are best right now. Gemini CLI is good but not as good. I setup a sandboxed Linux mini PC and installed Claude Code and OpenClaw with some throwaway accounts and it BLOWS my mind how good it is. Connected it to a burner WhatsApp account I can chat with to automate pretty much anything, including coding via chat message, rebuilding, commit/push to GitHub, etc etc. So much fun.
Crazy stuff. Yeah was talking to folks last night at a Xoogler meetup and OpenClaw was definitely one of the topics of the night. Definitely going to look at Claude Code. Seems to be where it's at.
Loved reading this. I feel like I either picked exactly the right time, or exactly the wrong time, to quit my software engineering job 18 months ago! I wonder if there's ever been such a rapid change in how computers are programmed before? My initial dabblings have followed a similar trajectory to yours: type a prompt into a generic chatbot and be amazed at the initial completely functional app that gets spit out; followed by uncertainty about the best way to refine / scale up from there. As you say, there are so many tools out there and new ones coming out each day, it's hard to know where to go... Hoping that reading this blog will give me the answer someday!
Simon I am now deep into Claude Code and it is going great. Looking forward to writing about my first impressions.
came for the pickleball data visualization and the accompanying story inspires us to all explore and experiment - thanks for sharing Brian :).
Thanks for checking it out. Yeah fun stuff. Definitely check out Johnkew Pickleball on YouTube if you want to go deep down the data driven pickleball paddle reviews. Fun stuff.
> The Churn: One Step Forward, Two Steps Back
Or as I put it: minutes 1-10: "Whee, I'm flying!" hours 1-5: "...into a tar pit of despair."
Figuring out how to work around tooling limitations is definitely one of the most frustrating parts of this for me, especially since it often seems that if you just wait another few months, the friction has been largely overcome for many of the tooling problems you spent days trying to develop hacks to get around.
But that magic "Wow!" for the first 10 minutes keeps luring me back, and I'm slowly improving my workflow to the point of it actually feeling like a productivity enhancement most of the time now.
That is good to hear from you Jessica. If you think it's net positive then it must be net positive as you are not one to sugar coat. I totally hear you on the wait a few months, and I'm kind of glad I sat out for 9 months in that sense. Things are changing so fast where you read many suggestions of adding this or that to your prompt then later reading that those workarounds are all unnecessary with the latest greatest models or tools. The workflows definitely feel like they continue to rapidly evolve.
I can't speak to the external toolchains at all, and I haven't been trying to do any UI work. For me, the big turning point was when I stopped trying to use it in an IDE and really dove in on the CLI front. Going the conversational route to see what the agent thought it should do tended to help me figure out mistakes in both its planning and my own thoughts on what needed to get done much more quickly. Encouraging it to suggest things and ask questions before we really got going on the coding front helped me better think through what I thought I wanted, and also seemed to help me frame things in ways that the agent seemed to grok.
I've had plenty of mini "Ah ha!" moments that have helped me start using the tooling better, but I suspect the majority of them will age as well as milk. Maybe after I return from Japan I'll have to wipe my brain of all my accelerating hacks, clear my own context window, and begin again!
On the whole, as my eyesight has been continuing to glissade down with no self-arresting, I was settling into a certainty that coding was just not going to be possible for me (obviously there are plenty of low-vision and blind programmers out there, but at some point the tooling hurdle feels so large, that moving into architecture design and management feels far more tractable). The ability to leverage agentic support for coding has reawakened my sense of confidence in being able to do actual code development as a via career path again. I'm actually submitting code that contains more than just markdown, GCL, or BCL these days, squeezed in between the TLMing!