As a pair, not a code generator. Alexandra sets the architecture and the data model, writes the scope for each change, lets Claude Code implement inside it, reviews every diff, and gates what ships with automated tests, a fixed evaluation set for the assistant, and per-call cost tracing. Every system on this site was built that way.
What stays with her: the architecture, the written scope, the review and the decision to ship. What she hands to the model: implementation inside a scoped step, refactors under an explicit plan, migrations, boilerplate, and a second opinion to argue against. The gates exist because AI-assisted failures tend to be silent: a strict lint configuration caught a hook-order bug, retrieval evaluation cases now cover a silent embedding wipe, and provider health checks caught a reranker that vanished with an expired key. The full write-up is at /how-i-work-with-ai.