Man, so much useful stuff here, and quite timely for my journey.
I finally jumped aboard the Claude Code train early this year and am still trying to figure out the right structure and process around getting the most out of it. I don't use it much for code, but there's lots of relevant stuff here for me as well.
I recall you being the person who recommended Obsidian to me a while back. At the time, I wasn't ready to figure out how to work with Markdown files and create those rich connections, but now I'm having Claude Code build out an ever-growing knowledge base for me in the Obsidian vault.
Are you using Obsidian for your Gemini CLI work as well?
Also: You talked briefly about curating and compacting context. Claude Code actually has this as a built-in feature: Every time you reach the "ceiling" of the token window, it automatically begins a compaction process where it reduces the chat to its essentials, saves all the outstanding actions, then "wipes" the context window so you can start filling it with new chat details without (hopefully) losing track of where you were. Does Gemini CLI have something similar? And have you ever tried/looked at the Claude Code implementation for it and how it compares to your own home-brewed approach?
I haven't tried Claude Code myself, but I have a few colleagues who do and it works incredibly well, but I think its more or less the same approach. You really have to customize it with your own prompts and skills and and commands to make it yours. Otherwise, for my taste, the model tends to be extremely eager to make changes and not to discuss those changes and to push back on the things you're trying to do.
Agreed. I noticed that as well. Very eager to just go without talking, to the extent where I sometimes have to interrupt it and give it additional feedback, etc.
Hey man, fun times indeed. First, yeah, Obsidian rocks. I'm still not entirely pleased with Obsidian setup but I'll share whatever I have in time. I think perhaps 50% of what's in here is helpful but there is a lot in this setup that's tailored for coding and I believe it would work badly on general prose. And missing workflows too, like finding interesting connections, etc., perhaps maintenance in Obsidian which would work differently.
Regarding context and Claude Code, yes, Gemini also has that. I think at this point these companies are copying each other and more or less converging on the same features, and models are pretty equivalent too, save for specific benchmarks, but it's more a matter of taste (and principles, perhaps) what you use at this point.
Yup. I am now in the process of working specifically with Obsidian because that makes my context agent-agnostic. This way, if I decide to switch to Gemini or whatever else, my setup doesn't change dramatically.
even without going that "far", on "simple" hardware issues in homebrew homelab exercises what once one had to trawl through forums to find is neatly found with some contextual parameters on how to adapt an Apple ppc G4 tower for a mini-itx board to host a truenas installation, of course this is a benefit that arises as the enshittifcation of old webpages like forums continues apace. Loading an 'old page' on a forum tends to suck the beejesus out of my browser to cope with all the @&#%@%@&&#
Thanks, happy to be of some help. Yeah, I know the feeling. Forums used to be the best part of the internet when I was a kid. Now we have Twitter, which also was fun for a while, but it's turned to a cesspool of idiots. Hope Substack remains somewhat sane for a while.
Man, so much useful stuff here, and quite timely for my journey.
I finally jumped aboard the Claude Code train early this year and am still trying to figure out the right structure and process around getting the most out of it. I don't use it much for code, but there's lots of relevant stuff here for me as well.
I recall you being the person who recommended Obsidian to me a while back. At the time, I wasn't ready to figure out how to work with Markdown files and create those rich connections, but now I'm having Claude Code build out an ever-growing knowledge base for me in the Obsidian vault.
Are you using Obsidian for your Gemini CLI work as well?
Also: You talked briefly about curating and compacting context. Claude Code actually has this as a built-in feature: Every time you reach the "ceiling" of the token window, it automatically begins a compaction process where it reduces the chat to its essentials, saves all the outstanding actions, then "wipes" the context window so you can start filling it with new chat details without (hopefully) losing track of where you were. Does Gemini CLI have something similar? And have you ever tried/looked at the Claude Code implementation for it and how it compares to your own home-brewed approach?
These are fun times indeed!
I haven't tried Claude Code myself, but I have a few colleagues who do and it works incredibly well, but I think its more or less the same approach. You really have to customize it with your own prompts and skills and and commands to make it yours. Otherwise, for my taste, the model tends to be extremely eager to make changes and not to discuss those changes and to push back on the things you're trying to do.
Agreed. I noticed that as well. Very eager to just go without talking, to the extent where I sometimes have to interrupt it and give it additional feedback, etc.
Hey man, fun times indeed. First, yeah, Obsidian rocks. I'm still not entirely pleased with Obsidian setup but I'll share whatever I have in time. I think perhaps 50% of what's in here is helpful but there is a lot in this setup that's tailored for coding and I believe it would work badly on general prose. And missing workflows too, like finding interesting connections, etc., perhaps maintenance in Obsidian which would work differently.
Regarding context and Claude Code, yes, Gemini also has that. I think at this point these companies are copying each other and more or less converging on the same features, and models are pretty equivalent too, save for specific benchmarks, but it's more a matter of taste (and principles, perhaps) what you use at this point.
Yup. I am now in the process of working specifically with Obsidian because that makes my context agent-agnostic. This way, if I decide to switch to Gemini or whatever else, my setup doesn't change dramatically.
thanks for this, it's very useful,
even without going that "far", on "simple" hardware issues in homebrew homelab exercises what once one had to trawl through forums to find is neatly found with some contextual parameters on how to adapt an Apple ppc G4 tower for a mini-itx board to host a truenas installation, of course this is a benefit that arises as the enshittifcation of old webpages like forums continues apace. Loading an 'old page' on a forum tends to suck the beejesus out of my browser to cope with all the @&#%@%@&&#
Thanks, happy to be of some help. Yeah, I know the feeling. Forums used to be the best part of the internet when I was a kid. Now we have Twitter, which also was fun for a while, but it's turned to a cesspool of idiots. Hope Substack remains somewhat sane for a while.
I also note you also are describing an "agentic engineering" approach [via] https://mkennedy.codes/posts/its-not-vibe-coding-agentic-engineering/
oops
https://addyosmani.com/blog/agentic-engineering/