How it evolved

Grok in a chat window

Feb 1 to Feb 10

Used: Grok on my phone, a terminal, my own two hands.

• Grok told me what to type. I typed every command myself.
• Installed Linux, Docker, my own password manager and file sync.
• Wrote a tiny homepage by hand.

Next: "i have gotten a few services going" (my first ever Claude prompt, Feb 11)

Claude Code in the terminal

Feb 11 to Feb 20

Used: Claude Code, on its entry plan.

• The agent types the commands and writes the code. I say what I want.
• Within a week the homepage was a dashboard, and my finance app existed.
• The entry plan ran out within the week. Upgraded.

Next: I wanted agents running without me sitting there.

A team of bots

Feb 21 to May 31

Used: OpenClaw bots on Grok and others, then OpenCode, then Codex.

• Named bots (coder, researcher, fact-checker) I messaged from chat.
• Scheduled security scans and daily reports.
• Mostly server work. Barely any app code.

Next: Nobody read the bots' reports, me included. Then I took a month off.

Back, and going ham

Jun 22 to Jul 6

Used: Claude Code in the terminal again.

• Built apps flat out: finance came back, the trader bot, FirePOS from zero.

Next: one terminal = one job at a time on one machine.

T3 Code: many threads, any machine

Jul 7 to Aug 27

Used: T3 Code on the desktop (and the phone app), Claude and Codex.

• Every task is its own thread. Many run at once.
• Heavy jobs run on the server while I keep working on the desktop.
• One shared context reaches every agent on every machine (Jul 29).
• Previews, reviews and a ship-it gate before anything goes live.

Next: my desktop itself was still stuck doing things the old way.

Omarchy: ask and you will receive

Aug 28 to now

Used: Omarchy Linux on the laptop, then the desktop (Sep 3).

• The desktop became something I just ask for: widgets, menus, shortcuts, game fixes.
• Each one lands the same day.
• More and more of my threads started on the desktop.

Next: typing was now the slowest part.

Voice

Sep 12 to now

Used: a mic toggle on one key, then Wispr Flow.

• Long instructions are spoken, not typed.
• An idea comes out as a rant, typos and all. That rant is the prompt.
• This page started as one of those rants.

Next: we'll see.

Local CI

Sep 20 onward

Used: a CI controller and disposable Linux workers on my own hardware.

• I needed repeatable checks across my apps without paying for hosted CI runs.
• The controller queues ready changes, tests them together and reports the result. Each enrolled app has its own checks, including browser tests where configured.
• Workers run with resource limits and leave logs and test artifacts behind. Their disposable disks are removed after the run.
• CI reports whether the tested changes passed. A separate guarded step handles merging, and going live still requires my approval.

How the CI checks and release gates work →