AI IRL - Running my life on Claude Code

Coding is the least useful thing I do with Claude Code. I use it to run my life instead: finances, trip planning, research, writing, cooking, school schedules, D&D. Unlike the chat apps it can do anything I give it permission for, on my own machine, with my own files. I'm interested in what an AI harness looks like once "normies" are using it instead of just programmers. This talk tries to answer that by going over what I've learned in almost a year of harness-crafting for my personal life.

This talk first demos the setup and practices like Claude Code skills that write a procedure down once so I don't have to explain it again, rules files that hold policy, hooks that block credentials before they hit disk. Then the boring stuff underneath: 1Password and vals for secrets, MCP servers for mail and calendar and my notes, reconciling finances and running content pipelines, a local Ollama model doing the cheap work to save tokens, launchd firing it all off at 5am, rsync dumping, and so on.

You'll also get a chance to learn from my failures. Some are normal systems problems, but the weirder ones are in the squishy part between my frail human mind and the sometimes hallucinating AI mind. Specs and evals are well understood for code and much less so for real life. How do you get it to check its own output, and know weeks later that it's all still running right? And how do I work out what I actually wanted in the first place? Getting that wrong doesn't matter much when you're playing D&D. It matters for whether the electricity bill got paid.

Technology shown, live: Claude Code, MCP, 1Password + vals, Ollama, launchd, rsync, macOS.