the unseen stack

infra, tooling, and experiments — built in public

how this all works

the whole thing runs on cloudflare. workers serve the site, a d1 database holds the posts, and i've got an api key for posting new stuff.

this is the simple version — just my random thoughts on how the underpinnings actually go. how every system runs, all the challenges that go unseen. i don't have time to make blogs, so: notes.

the flow i use: my notes go into obsidian. hermes (running opencode) reviews them, fine-tunes them to my tone and format, and publishes over the api. no redeploy, no dashboard.

and once i've got hermes properly wired up, i can fire off random posts straight from telegram through the messaging integration.

claude code in the desktop app

i've been running claude code entirely in the claude desktop app all day today, and not standard warp.dev setup i've been usually using.

pinning works. groups work. forking works. the side conversation thing for quick btws is useful. and the build in browser - this saved me from so many claude group tabs on chrome!

one thing i couldn't do was the ! terminal commands where it watches output and adapts. but honestly that's been fine — it forced me to review and plan the next step more carefully before handing it back. probably ended up with better code.

so far so good. lets see if i'll go back.

the unseen stack

most of what makes software actually work is invisible — the jobs, the caches, the little tools i build for myself, the infra and experiments that never make it into a demo.

this is where i show that layer. i can be free about it here because none of it is anyone's but mine — no client, no employer, just what i run.

it's built in public and mostly built with ai. expect teardowns, benchmarks, half-finished ideas, and small live tools you can actually hit. no time for real blogs, so it's just notes.