How I work with AI

For me, AI does two jobs. In planning it’s a rubber duck that talks back: I bounce ideas and goals off it until the plan holds up. In building it’s a resource distributor: I hand over the code and the tests, so my attention stays on the final design and the user experience.

It starts every session with standing instructions. These are the five I’d put my name to:

• Verify assumptions, including mine. A confident guess is still a guess, so a claim comes with the file, line or query behind it.
• Find the root cause before a workaround. Work out what changed before patching around it.
• Nothing is committed or posted without my go-ahead. It drafts commits, messages and docs, and I review and send them.
• Show me when there’s a better way than mine. I’d rather learn something new than assume I already know it.
• Passing tests aren’t the finish line. Walk through the user’s journey with fresh eyes, and make sure it looks right and is easy to understand.

I keep a heavy hand on design calls and creative work. AI can lay out options, and I make the call. With team communication and customer data I use it carefully: it drafts and checks, and I decide what goes out.

I overrule it when it’s wrong. On Fiscal Compass, its first pass ranked paying off a loan early below the other debts, on interest alone. I argued for the cash flow a year-early payoff frees, and the planner now measures that. Read the case study

Since April 2026, nearly 300 of my commits have been co-written with Claude. The bigger jobs start from a written brief: context, the request, success criteria and tests. I keep a template for them, and I’ve written about 30.

• When a teammate is stuck on a regression, I have it walk the history with me to find the change that caused it. That’s how I’ve traced a top-priority regression, a data-loss bug, and a tool bug dated against the commit log.
• A product manager couldn’t record a demo. I ruled out the previous day’s release with evidence first, then worked through the configuration with Claude, and had the root cause 19 minutes after the report: a test had changed a workflow setting on the shared test server.
• I helped write my team’s first shared guidelines for AI coding tools with the senior engineers, so engineers newer to it start with guardrails already in place.

Every case study ends its build notes with the same three-part box, so you can compare projects: what AI did, what I directed and owned, and one place I overrode it.