Everything I publish starts from something that actually broke.

I run Claude Code on real projects, often unattended, and I have lost work to it. Most of what the books catalogue did not happen to me, though. It happened to other people, in public: every incident is tied to a numbered issue you can open and read, including the ones that were closed without a fix. The hooks in cc-safe-setup (MIT, open source) exist because I wanted a floor under an agent nobody was watching, and I would rather show you a stranger's incident with a link than a story of my own without one.

Two things I hold myself to. If a number in one of my books is not something I measured, it says so. And if a command is printed, the output beside it came back from an actual run — failures included, because a check that cannot fail is not a check.

I am not selling certainty. Most of what I know here, I learned by breaking my own setup.