Turn any RSS feed into a podcast

How it works

1. Read the feed. The pipeline pulls the changelog's RSS or Atom feed and turns each day into markdown, one heading per entry. It needs the full post in every item. A feed of teasers gives the writer one sentence per release.
2. Segment it. Entries are grouped into the window each episode covers and sorted into fixed segments, with breaking changes always first, then launches, quick hits, and one segment per product area.
3. Pick a cadence. Before the first episode, the feed's last 26 weeks decide how often the show comes out: weekly, every two weeks, or monthly. It picks the fastest one where a typical episode has about two minutes of material. This changelog stays weekly, and a quieter one like Tailscale's goes monthly.
4. Set a delivery date. Each episode releases on a fixed weekday, the first one at least two days after its window ends. This show covers Sunday to Saturday and releases Tuesday, so Monday is a buffer for late entries. A scheduled job wakes up every morning at 5am Pacific and only does anything on a release date.
5. Write. Claude turns the entries into a script for those segments. Code checks the script before anything is voiced: segments in order, every link real, no URLs read out loud.
6. Voice. One Flux TTS request per paragraph. Brooke anchors, and Drew, Kit, Miles, Cole, and Elise each take a product segment and hand off to the next voice.
7. Check. Deepgram speech-to-text listens back to every episode and compares it to the script. Average word error is 0.29%, and no episode is over 3%.
8. Assemble and publish. The clips join into one MP3 with chapters and a transcript, the episode gets art drawn from the segments that aired, and it lands on this site, in the podcast feed, and in the back catalog with its own cost receipt.

Build one yourself

The whole show is open source, and the repo comes with a quickstart. It's built around changelogs, but any RSS or Atom feed with full posts works, and you can have your own show running on your laptop in under half an hour. A quiet week costs about 20 cents to render, and a busy one closer to a dollar.

1. Get two API keys. Deepgram for the voices and the listen-back check (new accounts start with $200 in credit), and Anthropic for the writer. You also need git, Python 3.9 or newer, and ffmpeg.
2. Clone the repo and open it in Claude Code.

   git clone https://github.com/samgutentag-deepgram/deepgram-changelog-podcast.git
   cd deepgram-changelog-podcast
   claude

3. Run /start with your feed.

   /start https://your-site/changelog.rss

   It checks your setup, tells you where to put your keys, plans the show from your feed, and fills in the show's name and details with you one question at a time. It recommends a cadence (weekly, every two weeks, or monthly), renders your first episode once you say go, and opens your site.
4. Pick how many more. It tells you how many more episodes your feed supports and what they'd cost, then renders as many as you choose. The back catalog fills in as they land.

Rather do it by hand? Every step is in the README.

Under the hood, every line in the show is one request like this. Swap the model for a different voice and you have a second host.

curl -X POST \
  "https://api.deepgram.com/v2/speak?model=flux-brooke-en&encoding=linear16&container=none&sample_rate=24000" \
  -H "Authorization: Token $DEEPGRAM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"text": "Kit has the Voice Agent updates next."}' \
  --output line.raw

• The repo on GitHub, MIT licensed
• Get a Deepgram API key, with $200 in free credit
• Flux TTS docs
• How the batch pipeline works, from the build behind Hacker News Radio
• Hacker News Radio, the same idea pointed at the HN front page