Healthy AI First · Episode 9 · 15 min

Practice 1, My Daily Briefing as a Podcast on Spotify

The first practice episode: I show you my AI daily briefing, which lands in my Spotify library as a private podcast every morning at six thirty. A local page collects my calendar, tasks and messages, Claude turns them into a text for the ear, and Studio by Spotify Labs (often searched as Spotify Studio Labs) voices it. Along the way: why an AI morning briefing for listening has to be written differently from one for reading, what went wrong when I voiced long texts, and how you can start small yourself.

Companion document

All studies, sources and claims from this episode in one PDF, to read along or look things up later.

Download the companion document (PDF)

Chapters

  1. 0:00Opening
  2. 1:01Welcome
  3. 1:53The Morning Before
  4. 2:48One Page Instead of Ten Tabs
  5. 3:56Rules Against Making Things Up
  6. 4:57Five Thirty
  7. 6:02Writing for the Ear
  8. 7:27A word from us
  9. 8:19Studio by Spotify Labs
  10. 9:10Six Thirty
  11. 10:32What Went Wrong
  12. 12:02What It Brings
  13. 12:56Try It Yourself
  14. 13:53Closing

Show notes

About the data

All data in the video (tasks, meetings, names) is made up; the setup is real.

Try it yourself

  • First: write down three questions. Which three questions should your morning answer? For example: What matters today? Who is waiting for me? What do I need to decide? A first briefing doesn't need more.
  • Second: let the AI write a text for the ear, not for the eye. Numbers as words, short sentences, fixed signposts. Listen to the first episodes carefully and check: is what you hear actually true?
  • Third: find your way into your podcast app. Studio by Spotify Labs is being rolled out step by step (Mac only, Premium only, 18+), and many people don't have it yet. The alternative: Save to Spotify, a small tool for coding agents like Claude Code or Codex, puts finished audio files privately into your own Spotify library (according to Spotify for Free and Premium users, as a beta with usage limits).
  • A practical note: my routine runs on a computer that is awake in the morning. Plan for that if your laptop sleeps at night.

How it works

  • A local page called "Today": a small web server on my Mac mini. It collects the calendar, tasks, the status of important tickets and messages that need an answer. It syncs tasks every 5 minutes, the calendar and inbox every 30 minutes. Everything stays local.
  • A scheduler from 05:30: it runs every 15 minutes and, from five thirty on, checks whether today's edition already exists. If something fails, it tries again fifteen minutes later. By 06:30 at the latest, everything has to be ready. Better no briefing than a wrong one: yesterday's text is never read out again.
  • Claude on the command line, without tools: the model has no access to any system, gets only the data as text and has to answer in a fixed JSON format with fixed fields (goals, decisions, meetings, radar, text for listening). This takes about 10 minutes; the checked edition is saved as a file that never changes again.
  • Times only from the calendar: a check after the AI compares every meeting time with the calendar and removes anything that isn't there. Without a calendar entry, the page says "time unconfirmed". Every statement carries a label: fact, suggestion or open.
  • A spoken-text program after the AI: it writes times, amounts and numbers as words, gives abbreviations a spoken form, takes the pronunciation of product names from a pronunciation dictionary I maintain myself, and removes brackets and special characters. At the end, no digit and no special character may be left.
  • Paragraphs of 40 to 70 words: each one ends at the end of a sentence; a small algorithm works out the best split. On top of that, a fixed order with signposts and a short summary at the end.
  • Handoff to Studio through one folder: Studio may only read its own working folder and gets only the finished text. My server mirrors it there as a text file every 2 seconds; the first line holds the date, the edition and a status (ready, not ready, already done).
  • A routine in Studio at 06:30: it checks the first line, creates a private episode in a separate show called Daily Briefing, paragraph by paragraph, and then leaves a small receipt file in the same folder. The server records "episode created" and archives the file, so no episode is ever created twice.
  • Whisper check: I now generate long manuscripts paragraph by paragraph with the command line tool for speech output that comes with the Studio app, without the assistant in between. Every paragraph is turned back into text right away with Whisper (running locally) and compared with the script; below 80 percent match, it is generated again. Result in my case: umlaut errors down from almost half to under 1 percent, and from more than 5 hours to about 20 minutes.

Sources

The Morning Before

Writing for the Ear

Studio by Spotify Labs

Save to Spotify

What Went Wrong

  • Whisper: Radford, A. et al. (2022): Robust Speech Recognition via Large-Scale Weak Supervision. The speech recognition model I use locally to turn every paragraph back into text. https://arxiv.org/abs/2212.04356

What It Brings

Healthy AI First is AI-researched and AI-generated, narrated in an AI version of my voice and reviewed by me (Björn).

Bringing AI into your team?

Björn is one of the founders of CLN Work. CLN Work helps companies adopt AI in small, verifiable steps: finding where real capacity is, which processes can be automated, and how to use AI interactively. The decisions stay with you.

Book a free 30-minute conversation