Refactoring Orgs

AI is hiding more and more code from developers. But there's a counterintuitive pattern: the developers reporting the best results with AI assistance are the ones who became more explicit about structure, not less. They specify file responsibilities upfront. They write design documents before any implementation. They give the agent named tasks with known locations and defined boundaries.

This is the opposite of what "vibe coding" implies. The effective pattern is not: describe vaguely, accept whatever arrives. It is: specify precisely everything except the syntax. These developers extracted the mechanical part of programming - translating a known structure into correct syntax - and kept the architectural reasoning for themselves.

The developers failing with AI agents are the ones delegating architectural decisions alongside syntax. That is where catastrophic problems appear: functions that instantiate new interpreters on every call and still pass tests, duplicate implementations with identical names, tests rewritten to accommodate broken code instead of implementations fixed. An agent cannot reason about architecture from a vague prompt. It fills gaps with plausible-looking structures that are quietly wrong.

The practical implication is uncomfortable. AI assistance does not lower the bar for software engineering. It raises the value of one specific skill: thinking through a system structure completely before writing any of it. Developers who already did this habitually got a power tool. Developers who relied on emergent structure while coding got a mirror showing exactly where their process breaks.