Onboarding a Teammate with Amnesia — context engineering with Claude Code on a production banking app

Your coding AI works like a brilliant contractor with amnesia: every session starts from zero, and the bottleneck is usually the input, not the model. This talk shows how an iOS team turned Claude Code's context into an engineered artifact on a production banking app. The context is layered, from org policy to project rules to session memory, and maintained like code: new rules need evidence, decisions and rejections get written down, stale facts get pruned.

The proof is a UIKit-to-SwiftUI navigation migration across a dozen modules. The first one was slow and full of mistakes. Each mistake became a durable rule, and within a few modules the bottleneck had moved from AI output to human code review. The examples come directly from the project: our context files, the corrections we made, and failures, like the missing line of context behind a git reset. I'll show what worked, what didn't, and what survived real production use. None of it is specific to Claude Code.