How I use AI, with guardrails

I use Claude for most of my day-to-day work. That only works because of the guardrails around it.

• In products: structured output in Laravel (my AI import tool maps messy CSVs with typed responses), RAG over notes and tasks, and vision and speech to text where they help.
• In my workflow: Claude Code every day, MCP servers so agents can use my tools safely, and Lara, my own harness that runs agents on tickets and then asks me.
• Guardrails: changes are proposed, reviewed, then applied; agents work on branches, not production; every tool call is logged, and cheaper models do the simple steps.

The short version is on my AI policy page.