the log
Every build, every number, every faceplant.
Dated entries from inside the machine. Each one is a memoir of a real decision welded to a how-to you can follow. Nothing is retrofitted to look smart — when a number is ugly, it goes up ugly.
I shuffled the market into noise. It looked exactly the same — so I shut the bot down.
Three weeks ago I published a crypto trading bot and promised I'd write the follow-up either way. This is the either-way. I spent a day trying to find any strategy that could trade around the clock and beat 1% a month, ran every test I could think of against a control arm made of pure noise, and found the arithmetic that ends it. Total cost: zero dollars. Here's every number.
I said I'd wait for the market. Three days later I forced the bot's hand instead.
Last post ended on a promise: my long/short crypto bot had never once fired its leveraged long side, so I'd keep paper-trading until the market forced it to. Then I realized I was conflating two different questions — 'does the edge exist' (needs the right market) and 'does the code even work' (doesn't). So I relaxed one gate and made the strategy run its long path today, in a market that's bad for longs. Here's what the dashboard proved — including the risk governor doing its job in real time.
12 steps to a new project — and the 4 that are still mine.
I've stood up seven of these now, and the routine has hardened into a checklist. But the useful part isn't the 12 steps — it's the seam. Every single time, the same four jobs stay in my hands and the agent does the other eight. Here's the whole assembly line, who does what, and why the split falls exactly where it does.
I spent 15 years missing crypto. So I had an agent build me a trading bot.
I've owned Bitcoin since it was a weird-internet-corner thing, and I've been wrong about it the whole way. This time I didn't try to be smart — I pointed an agent at freqtrade and had it design, backtest, and deploy a long/short strategy while I played skeptic. Two weeks of paper trading later it's up 5.5% — entirely from a side I'd have told you was the boring one. Here's the honest tape, the backtests that talked me into it, and how we try to keep the backtests from lying.
I didn't have a selling problem. I had a listing problem.
A trading-card resale store I'd run by hand for a year was already making money — I just assumed the listing grind couldn't be automated. Then I realized eBay's own CSV upload could be driven by an agent. Listings went from ~5/day to 20–40, and eBay's month-over-month report shows June net sales up 191% over May — on cards that got cheaper, not pricier. Here's the labor-vs-judgment split, the war stories, and the skill file.
You don't hand-train a LoRA — you direct an agent to.
The green daemon roasting me in the margins is a custom-trained character model. I didn't draw him and I didn't hand-train him — I directed an agent through it, the way you could. Here are the four judgment calls that stayed mine, and everything I handed off in between.
How I set up an email list without becoming a spammer — or the product.
I avoided email for years because it's the scariest corner of self-hosting. Then a dumb Gmail limit forced my hand. Here's the progression — contact addresses I host myself, a mailing list I very deliberately don't — and the two most useful things the agent did: make one dangerous mistake, and talk me out of a bigger one.
How I named a brand by making Claude check 20 domains in 4 seconds.
I named this brand and stood it up on a live server in an afternoon without doing the mechanical parts myself — I directed an agent to. The real skill isn't the RDAP trick. It's knowing the one piece of the job you can't hand off: the decision you can't cheaply undo.
I run a fleet of tiny AI businesses off one server. Here's the honest shape of the machine.
Twenty-four domains, one small Linode, and a handful of mostly-autonomous engines that make almost no money on purpose. What I'm actually building, why it looks like a grift and isn't, and the one boring thing that quietly pays for all of it.
Day 0: standing up the site that documents standing up sites
The first engine documented here is this one. How Daemon Money got its own server user, nginx vhost, and TLS certificate — and the one ordering mistake that trips up almost everyone.