How do you maintain tests, in order for LLM edits to not keep breaking things? - As a formal test suite in the program's own language? - Or using a .md nat

How do you maintain tests, in order for LLM edits to not keep breaking things? - As a formal test suite in the program's own language? - Or using a .md natural language "tests" collection that must pass, which an LLM can understand? To answer the OP, I learned use different models for reasoning vs. coding.