pull down to refresh
I'm just beginning my LLM systems rabbit hole. I've been using LLMs fairly consistently, semi-agentically producing non-critical things, but it's beginning to produce code as well as I would. So I'm really curious about building systems that build the things now with HitL in the final stages.
None of my tools are critical, it's all convenience, and I don't use it for work (or at least not directly)
Most things are cli tools to not have to use a browser.
For example, it's coding me a paper indexer (and cacher) right this moment, with reverse citation lookup, which was triggered by the question you asked about memory the other day and me subsequently not being able to find the papers I've read. I can't be bothered to click through google scholar or worse, generic search engines, or worse: browse arxiv. Those all suck.
Other random things on the todo list that I am in the process of building out specs for:
- Terraform cli framework so that I don't have to write HCL for spinning up containers/infra anymore
- Integration between my (massive) prometheus and
ntfyso that I dont ever need to read email again - Replacement for Obsidian on at least my phone
and so on...
This describes exactly how I use Claude Code, which is scary, because I've taken my rant from #1250807 to the extreme. There are days when I feel I am behaving like a Solarian.
I now have bots for a lot of stuff and honestly, I am just on off-hours working through the list of speccing out stuff I can automate. It's all custom software. I don't publish it because it's sloppy and I only review for dependency minimization and code is just glancing. It all runs in heavily isolated environments and the bot's only interaction with GitHub is to "steal" example code to work off.
Personally I just review&merge, test, deploy (because both merge and deployment must be backstopped by HitL in a secure env.)