One dev, many projects, built in the open with AI. Honest notes from the daily work.
The architecture of my order system: how work enters from emails, sessions and error reports, how Claude Code decides what runs unattended in a git worktree, and the cost caps and prompt injection checks that keep it from going wrong.
How a single TODO.md per project grew into an order queue across ten projects, and why every safety gate in it exists because something broke first. Claude Code multi-project orchestration, honestly measured.
I hit the Claude Code weekly usage limit with four days left in the billing cycle, no warning at all, and had to buy emergency tokens at list price. Here's how I built a usage meter that actually tracks the weekly limit, in four attempts.
I gave ten projects one order queue that runs itself every five minutes. Here is how it is built, the rules that keep it safe, and the night it buried me in its own status reports.
Three parallel Claude Code sessions, one shared config file, no lock. One deleted an exception another had just added and took a production feature offline. The fix was not a better prompt, it was a hook that returns an error.
My content pipeline never researched a single number. A mandatory statistic field quietly forced the model to invent them, with real-sounding sources, in content people rely on. The fix was not a better prompt. It was deleting the field.
Sixteen months of fully automated content, one honest look at the Search Console data, and the moment I killed the machine. The mistake that made me start writing by hand instead.