I burned a ton of tokens this summer trying to document our legacy codebase in hopes of quantifying parts ahead of a refactor. My conclusion was that LLMs are bad at this. It waffled between unhelpfully verbose to omitting key aspects. I had to manually review each page. It really struggled with cross file references and inheritance. I tried several approaches, top down, bottom up, text first, diagram first. Maybe I'm not the prompt wizard I need to be. But I would never trust AI summary of any code longer than 500 lines.