From recording to release — how AI works with me on the podcast

I sometimes get the question: "Do you use AI on your podcast?" The answer is yes — but probably not the way people imagine.

The AI doesn't conduct the interviews. It doesn't write the questions. It doesn't sit across from the guest and listen. But once the microphone is off — that's when it goes to work. In this episode I sat down with Dr. Sigrún Stefánsdóttir and walked through the workflow openly, because I think transparency matters with this technology, and because the process itself shows both what AI can do and where its limits are.

The conversation is always the raw material

Before anything happens on the computer, something has to happen between people. The conversation is the source element — the voice, the context, the asides that turn everything around. None of that is born in an algorithm. The AI works from the conversation; it never replaces it.

The workflow

The recording is captured at high audio quality — the foundation everything else rests on. It goes through speech-to-text (for Icelandic I use models trained on Icelandic speech and Google Cloud Speech-to-Text, plus my own open tool, whisperSSTis, an open-source Whisper model tuned for Icelandic). The output is a working draft — not perfect, but a good enough base, which I correct for obvious errors, names and key terms.

Then the interesting part: the transcript goes into a custom system prompt and the AI returns two kinds of work — an episode analysis (quality, key topics, notable lines, a directory-ready summary, and a first blog draft) and a creative pass (a visual description for cover art, music suggestions, and prompts for image and music models). It's like a very organised, hard-working assistant that reads everything and proposes ideas. But — and this is the key — the depth is still missing.

So I review everything the AI returns. Verify the quotes. Fix tone and style. Add the context only the person who sat in the conversation can give. The AI proposes; I decide. From there comes the promotional material — cover image, audiogram clips, a musical mood for the episode — and a final holistic pass before anything is published: is the blog in the right tone, are the quotes correct, does the promo actually reflect what happened in the conversation? That's the human gate, and it's inseparable.

Why it matters

The goal is simple: speed up the process and improve the quality without sacrificing human judgement. Before this workflow, most of my time went into what the AI now handles in seconds — transcription, structure, first drafts. Now I can spend that time on what actually matters: reviewing, reflecting, and thinking about what the conversation means.

Show notes written up from the conversation with AI assistance and edited in my own voice. The raw transcript is available separately — the soul stone, for those who want to dig deeper.