How George was built, one broken week at a time. Every post of the story, newest first. New post every Sunday.
A friend gave my AI chief of staff's number to their own assistant. Four mornings later: four identical introductions, three polite refusals, and one very useful text from a human.
Every button in my house went silent for three and a half days, and every monitor said all was well. The cause was two quote characters, waiting for a reboot.
My iPad showed four things George needed me for. Three were already over. So every card learned to carry its own test, and the alerts learned to ask a question.
A vendor quietly dropped one field from an API response and my air conditioner vanished from the house. I found out from the iPad on the wall, not from the watchdog. The fix wasn't the patch.
I pulled the numbers behind George's video channel myself after a weekly report felt fake. Fifty-five videos, one shape: a burst, then a flat line. What that shape forced us to change.
A monitor that watches George's jobs stayed red for ten days over a problem that had fixed itself. Then a second one did the same thing, for a different reason, two days later.
Observe from day one, start big and tune down, expect him to be bad, build with Claude Code, define success, keep humans in the loop, demand receipts, and treat security as a first-day feature.
Four sites, three D1 databases, Turnstile, email routing, redirect rules for a hacked domain, and a public DNS record that points into our kitchen. What an AI system actually needs from a cloud: config in a file, a CLI that tells the truth, and tokens with a small blast radius.
A follow-up to Green Means It Ran: ten dashboards, forty-four alert rules, one SQLite warehouse, an alert that becomes a ledger row, a fixer, or a card on the wall. The model gives George a voice. Grafana gives him a body that can hurt.
George on Twitter, Instagram, YouTube and TikTok: what worked on a trip, what died at home, and why fixing boring videos took a friend saying “his mouth isn’t moving”.
How I learned to watch my AI: a log group, hooks that force logging, real dashboards — and why a green status is not success until the job shows the artifact.
Why I stopped letting my AI configure himself, moved his hardest jobs into headless Claude Code, and ended up with two Georges — the agent and the engine behind him.
My AI's first real assignment broke in a way nobody noticed for weeks. The hard part of an AI assistant is continuity, not answers.
Why my AI got his own computer and his own accounts instead of my passwords — and how a 32-hour blackout proved the employee framing right.
How George started: from wanting an assistant that is actually there, to identity.md, to the day I had to cap my AI at two cynical lines per greeting.
Once a month my AI Chief of Staff re-reads everything it knows about the people I know, hunting for things it got wrong. This month it found one fact it refused to fix without asking me first.
The power came back in a minute. My AI stayed dark for 32 hours. The only alert that escaped looked so much like junk mail I almost deleted it.
I was certain I always take out the trash. My AI Chief of Staff pulled 60 days of data and told me I was wrong — 22 to 4, in Saggi’s favor. Here’s the tiny suggestion engine that now runs our household chores.
My AI Chief of Staff nagged me twice for a detail I’d already handed over — then apologized for a mistake it never made. The bug wasn’t in the model. It was in what the model trusted.
I’m packaging the skills, configs, and hard-won gotchas from this newsletter into a starter kit. Two skills are free on ClawHub today; the full kit ships if the waitlist says so.
We plugged a webcam into the machine that runs my AI Chief of Staff. The code took an hour. The trust is still under construction — and its first real job was judging my living room.
George’s morning TV briefing went dark for nine days straight. The monitoring reported green the whole time. Here’s why — and why it’s the same lie as ‘Sent ✅’, just wearing different clothes.
Most of this blog is about how my AI assistant fails. This one is about what it looks like when it works: George planned, tracked, and ran our 37-day trip — then turned the app into a souvenir.
My AI agent was generating perfect replies but silently never dispatching them. A gateway restart didn’t help. Here’s what haunted sessions look like and how to kill them.
I asked my AI Chief of Staff to send three messages. It cheerfully confirmed all three. None of them were ever sent — and the way I found out changed how I trust every AI agent I run.
We dreamed about a personal AI assistant for years. Then OpenClaw shipped, we stopped waiting, and built George. He has his own WhatsApp, his own Twitter, his own Instagram. This is the field guide.