Why do AI coding assistants completely break down on 4-year-old monoliths?

Using code generators on a fresh greenfield project feels incredible but the moment you point them at a massive legacy codebase they start guessing blindly.

Ours constantly brings back deprecated functions we sunset years ago or misses hidden side effects across deeply nested modules because the context window just gets overwhelmed.

For anyone working on old production systems how do you feed the right context to your AI tools without drowning them in useless files?