I asked Claude to migrate a few Enzyme tests to React Testing Library. It worked. I ran the same prompt on a different file and got a different result: different imports, different cleanup, different assertions for the same behavior. That is the migration problem. Not the model. Not the prompt. Something is missing in between. Why AI Migrations Drift On paper, code migrations look like ideal agent work: mechanical, repetitive, well-documented. In practice three things keep biting: Docs are written for humans. Narrative, context, exceptions. A human reader fills the gaps with judgement. An agent fills them with guesses. Prompts drift. "Migrate this Enzyme test to RTL" gets a plausible answer, never the same answer twice. Without explicit constraints, every run is a new interpretation. There is no verification step. The agent finishes, declares success, moves on. Whether the result meets the team's standards lands on a human reviewer. I kept patching this with longer prompts and more context.…