How to Use AI for Podcast Show Notes and Chapter Markers (Without the Generic Output)

Podcast show notes used to be the worst chore in the whole production pipeline. You finish recording, edit the audio, push it through your hosting platform, and then sit down to write a summary that nobody reads but every directory wants. Most podcasters either skipped them or wrote three rushed sentences that did nothing for discoverability.

AI fixed half of this problem and made the other half worse. Drop any transcript into ChatGPT and you'll get show notes back in twelve seconds. They'll also sound like everyone else's show notes. Soft language, padded summaries, the same bullet-point structure as the next twelve podcasts in your category. Discoverability through sameness is not a strategy.

This is the workflow I actually use, with the parts that matter.

Start with a clean transcript, not the AI tool's transcript

The single biggest quality jump comes before you involve AI at all. Most podcast platforms now ship a "free" AI transcript along with your audio file. Some are fine. Most are bad enough that the show notes get poisoned at step one — misspelled guest names, dropped technical terms, missing speaker labels.

Pay the small amount for a real one. Whisper-large (free if you have the GPU, or via OpenAI for around $0.006 per minute) is still the cleanest. Deepgram Nova-3 is faster and decent for clean studio audio. Descript bundles a transcript with its editing workflow and is a defensible choice if you live in Descript anyway. Avoid anything that uses a generic model trained on call-center audio. It will mangle every proper noun you care about.

If your transcript has the wrong guest name in it, your show notes will have the wrong guest name in it. AI does not catch this. You do.

Use a prompt that fights the generic-podcast voice

The default ChatGPT prompt — "write show notes for this podcast episode" — produces the exact slop you're trying to avoid. The model has been trained on thousands of bland Spotify summaries and that is what regression to the mean gives you.

A prompt I have had luck with:

You're writing show notes for a podcast called [SHOW]. The audience is [SPECIFIC AUDIENCE]. Read the transcript below and write: (1) a 2-sentence opener that names the most interesting argument the guest made, not a vague topic teaser; (2) 4–6 specific takeaways, each phrased as a claim or recommendation, not a vague "we discussed"; (3) any books, tools, or people referenced, with timestamps if possible. Do not write "in this episode" or "join us as we explore". No emojis.

That last sentence does more work than the rest of the prompt combined. The model loves "in this episode" and "join us as we explore" because every podcast in its training set opens that way. Banning specific phrases is more effective than asking for "natural" or "engaging" language.

Chapter markers are easier than people think

Spotify and Apple both support chapter markers now, and they meaningfully improve completion rates for long episodes. Generating them used to mean either listening back with a stopwatch or paying for Castmagic.

The simpler path: feed your transcript to Claude or GPT-4 with timestamps included and ask for chapter markers every 5–8 minutes, with titles under 50 characters. Claude tends to be better at this because it is less inclined to invent timestamps that aren't in the source. GPT-4 will sometimes make them up. Spot check anything before you ship it.

If you want it fully automated, Snipd, Swell AI, and Castmagic all produce decent chapter output. Snipd in particular tends to pick natural transition points rather than fixed intervals, which sounds nicer.

Always check three things before you publish

The mistakes I see most often in AI-generated show notes:

• Hallucinated quotes. Models will paraphrase a guest in quotation marks. If the words aren't verbatim in the transcript, they shouldn't be in quotes. Search the transcript for any quoted line in the show notes.
• Wrong attribution. When two guests speak, AI frequently swaps who said what. Especially bad with unmoderated panel formats. Speaker-diarized transcripts help; manual review is still required.
• Missing the actual interesting thing. Models gravitate to the abstract topic and skip the specific moment. If your episode contained a story about a guest losing a million dollars on a bad bet, that's the show notes opener, not "We talked about risk management."

The tools worth paying for

If you record one episode a week and don't want to think about any of this, Castmagic is the path of least resistance. About $30/month for show notes, chapter markers, transcript, and social clips. The output is fine. Not great, but fine, and that is a fair trade for forty minutes saved per episode.

If you record more, or you care about the editorial quality, build the pipeline yourself: a real transcript service plus Claude or GPT-4 with a sharp prompt. The marginal quality is worth it, and you'll spend maybe twenty minutes per episode instead of an hour.

What you should not do is the worst of both worlds: use whatever your hosting platform throws in for free and publish it untouched. Your listeners may not notice, but every other podcast in your category is doing exactly the same thing, and the algorithm sees the sameness.

Tomorrow: chapter markers and clip pulls for video podcasts, where the AI tooling is moving faster than the audio side.