Podcast Show Notes Generator: Transcript to Notes in Minutes

A podcast show notes generator doesn't have to be another $29/month SaaS. Mine is one API call plus one prompt, and it produces the summary, topic list, verbatim pull quotes, and resource links for an episode in about ten minutes — including the fact-check pass. I run it for two shows; here's the whole thing.

Episode to transcript in one request

If your episodes go up on YouTube (most video podcasts do), the transcript is one call away:

curl -s "https://youtube2text.org/api/transcribe?url=https://www.youtube.com/watch?v=EPISODE_ID&maxChars=150000" \
  -H "Authorization: Bearer yt_your_key"

A 40-minute episode comes back as 45-60k characters; a 90-minute long-form conversation runs 100-130k, still inside the 150000 maxChars ceiling and inside a single Claude or GPT context. The response includes a truncated flag, so you'll know if a marathon episode got cut rather than discovering it when the notes stop at minute 70.

The prompt that keeps quotes honest

Pull quotes are where generic summarization prompts fail you. LLMs love to "improve" quotes — tighten the grammar, swap a word — and a paraphrase inside quotation marks is how you end up getting a correction email from your guest. This prompt forces verbatim extraction:

You are writing show notes from a podcast transcript.

Episode: {EPISODE TITLE}
Show: {PODCAST NAME}
Guest: {CORRECT GUEST NAME + one-line bio}

Produce:
1. A two-sentence episode summary. No hype words.
2. 5-8 bullet topics in the order they were discussed.
3. Three pull quotes copied VERBATIM from the transcript.
   Do not paraphrase. Do not fix grammar. If a quote needs
   context, add it in brackets before the quote, never inside.
4. Every book, tool, person, and company mentioned, as a
   plain list for the resources section.

If no quote is strong enough, say so. Never fabricate one.

Transcript:
{PASTE TRANSCRIPT}

Note the guest name goes in as an input. That's deliberate, and it's the fix for the next section.

The guest-name problem

Here's the honest caveat: these transcripts come from YouTube captions, and auto-generated captions mangle names with total confidence. I've watched "Swyx" become "swicks" and a guest named Priyanka spend an entire episode as "Brianca." Feeding the correct spelling into the prompt lets the model normalize every mention in the summary and topic list. But the verbatim quotes need one more check — if the caption already garbled the sentence, "verbatim" faithfully reproduces the garbling. Before publishing, play the section of the episode each quote came from and confirm the words. That's the fact-check pass, it takes three or four minutes, and skipping it is how mangled quotes end up screenshot on someone's feed.

Slotting it into a publishing workflow

The manual version above is fine for a weekly show. Past that, two upgrades are worth it. If you want notes generated automatically every time an episode publishes, the n8n workflow does the whole chain without code. If the episode deserves a full article rather than notes, the same transcript feeds the video-to-blog-post pipeline — and if you email your listeners, those per-episode summaries stack directly into a channel newsletter.

The shared demo key covers 5 episodes a month, which is one show's worth. For the full parameter list, error codes, and both auth header formats, the API reference lives at youtube2text.org/api.md — the whole thing fits on one page, which is roughly the energy this entire workflow runs on.