gh_TechLuddite describes moving a Road to Vostok game-playing AI from cloud JevJev to local Jeff on his own PC, training with LoRA from his play, and credits Omarchy and Claude Code in the outer improvement loop.
nimrodnnagy says "Omarchy + Hermes + T3Code is GOD MODE." and explains "Action button on IPhone triggers voice recording which is sent to Hermes which starts a new T3 thread." So they can say "Hey, add rate limiting to the login of project X" and off it goes.
ThePrimeagen's Omarchy QA update says there are two things that are very hard about building a QA agent harness: how to go fast in the day and age of agents (which are very very slow), and "What is correct? What do we even test?"; a basic agent crawling the desktop in QEMU originally took ~15 minutes to open the lock screen, type in the password, and confirm "that we are back on the desktop", a few changes got that to ~5 for median, and with more changes and a custom harness it is down to ~2 minutes, which "still sucks", so for the last week ThePrimeagen has been experimenting with Clef ("cloudflare's jev, it has image capabilities, which is what i needed"), hoping to get logging in, system menu, lock screen, password, back to desktop, and confirming it is done down from 2 minutes to 30 seconds; if that works, ThePrimeagen can officially move to step 2, which they are not stoked about and may need some help with, and they ask anyone interested in helping to let them know.
Dennis_Rye says he improved GPT-6.1 Sol review performance by about 40% on Omarchy PRs through prompt tuning, that Opus 5.5 is still better but this brings Sol to 90-99% while costing around 30% as much, and links a prompt write-up; the chart compares a rising tuned-prompt line against a flat untuned GPT 6.1 baseline across v1 through v10.
