We took over a legacy codebase with a one-month handover, no tests, unclear architecture, and a mandate to use AI for everything. Some of our stakeholders just expected magic, but here's the catch: AI is most useful when you already understand the domain well enough to assess its output. And we didn't. Not yet. We were just learning the domain, the technology, and AI tooling all at the same time, under pressure.

This talk is about making the implicit explicit, step by step. And in this system everything was implicit: the domain knowledge, the design decisions, the boundaries. We generated documentation to evaluate the current implementation and to create a shared picture, modelled the domain together, wrote assumption tests to understand what the code actually does and created instruction files, skills and agents to define guardrails for the AI (or Harness Engineering as it's called nowadays).

We learned that AI doesn't replace domain understanding, it amplifies it in both directions. It needs context it can't build on its own. When we knew what we were doing, it was a multiplier. When we didn't, it confidently led us astray. And yes, our AI setup also broke production along the way.

But also on a personal level we faced some challenges, not everyone in the team was convinced by the hype and some maybe even too much. It forced us to have some honest conversations about what AI can and can't do and to align our expectations.