AI policy

I use Claude for most of my day-to-day work: planning, writing code, reviewing and testing. This is how I keep it safe.

• I review everything it produces. AI writes a lot of the first draft; nothing merges until I’ve read it.
• Tests run before anything merges. Agents work on branches and in sandboxes, never directly on production.
• Real data needs a person. Anything that changes customer data is proposed first, then approved by a human.
• Production is dry-run or read-only first. Scripts show what they would do before they do it.
• Everything is logged. Every tool call an agent makes is recorded, along with what it cost.
• Secrets stay out. Passwords and keys are never sent to a model or indexed for search.

Posts on this site are written by me. Some are drafted with AI help and always edited by me before they go up.