It is important to be clear on how this podcasting stuff happens, it all takes inordinate amount of time and I put a significant amount of the time I can spend into preparing for that,

The show notes are accurate but a number of technologies are used to speed this process up

1. Write draft notes using word processor
   • References include Books/Papers and anecdotes
2. Take this, and record using Logic Pro
3. This recording then gets Transcribed using a Large Language Transcription model
   • Software: Whisper
   • Current Model: Parakeet from Nvidia
4. This produces a wall of awful text – but with time stamps.
   • This gets parsed through Claude’s Sonnet 4.5 using a prompt to take many fragments into ordered text,
   • It provides a changelog, but the prompt bars it from change, its only allowed to re-arrange from ~250 fragments
5. The Podcast itself, is cut/chopped/edited, ‘mastered’, uploaded to hosting.
6. The banner images are all AI generated weirdness, a mix of Apples ‘playground’, Adobe AI, and more recently Sora
   • Its an image – not a masterpiece and hopefully conveys something of the episodes nature.
7. The show note production, a significant time burden
   • originally all written by me,
   • Then written by me and handed to GPT to tidy up
   • Then written by me, handed to Claude Sonnet + given transcript to ensure alignment with podcast
   • And then full circle as the amount of work making sure these models didn’t add their own spurious material became too annoying as going and cross checking it against my notes became more of a time sink than writing it.
   • Now writing notes straight into the website, and trying to cut out the iterative process – and I’ll see how that works.
8. The SEO schemas (you cant see) java script are produced with Sonnet 4.5 and parsed with google rich results tool for validity
9. For Social media/marketing copy I lean more heavily on Claude because again, time burden in life.