Twenty years, two disciplines

My main tool is Claude Code. I also work with OpenAI's Codex. But the tool matters less than what I put around it.

I build what I call an AI operating system for a project. That is a set of files, rules, and checks that tell the agents what we are building, how it has to be done, and how they will know when they are wrong. Skip that part and an AI writes code that looks right and misses the point. Put it in place and the same model produces work I will put my name on.

The work gets divided. One agent plans the shape of it. Another writes. Another reviews that writing against the brief and is allowed to reject it. Smaller sub-agents take single jobs, like checking a page's markup or chasing one bug, then hand the answer back instead of everything talking at once.

Running more than one system matters too. Claude Code and Codex make different mistakes. When both reach the same answer on their own, that answer is usually solid. When they disagree, that is the part worth looking at closely.

I use this to build applications, websites, content systems, and search architecture — both the classic kind that ranks in Google and the newer kind written to be quoted by answer engines and AI results.

The obvious worry with any of this is that an AI simply makes things up. So I built Claude Watch: a private system that keeps a curated wiki of sources I picked by hand and feeds it to the agents, so they read verified facts instead of guessing from memory.