I wanted to know if an LLM could generate a podcast I'd actually choose to listen to — turning Hacker News threads into audio, in my case. Turns out writing the script is the easy 20%. The rest is fighting everything the model wants to do by default. Two things made the biggest difference: 1. Constraints beat instructions. Telling a model "be conversational" does nothing. What worked was giving the two hosts different information — one only read the article, the other only re...