One Evening to Build It: Article → Podcast on Apple Podcasts

How many "read later" tabs do you have open right now? You don't have to answer. I know the feeling.

Over the years I tried everything. Feedly, Obsidian, "I'll email it to myself." None of it really worked, because the problem wasn't the collecting tool. The problem was that reading articles competes with everything else.

So I stopped collecting. I started listening.

Email as the interface

I set up a dedicated mailbox that acts as an entry point for all my automations. My AI agents have full access to it — it's their "inbox," my way of communicating with them instead of Discord or WhatsApp.

The whole intake flow looks like this: I send a URL to that address and forget about it. Every evening launchd on my Mac fires a script that checks for unread messages. If it finds a link from my address, the pipeline kicks off. No app, no webhook, no clicking. Just email.

Bonus: newsletters as a source

I have a lot of subscriptions landing on that same mailbox. The system recognises them. Instead of treating a whole newsletter as an article, an LLM extracts all the links to individual pieces and saves them as tiles in my private widget.

There I can see the title, a short description, and the source for each link. I mark which ones I want turned into a podcast, which were valuable, and which to skip in the future. Instead of reading a newsletter, I scan a list and pick.

My AI agents drop their own finds into the same widget.Crystal and the rest of my agentsbrowse the web every morning looking for things that might interest me: Claude release notes, new libraries, articles outside my usual information bubble. They arrive via MCP, which acts as a shared inbox. One widget, many sources.

Here's the full flow:

A podcast in Apple Podcasts, next to my favourites

Google's NotebookLM can turn any article into a podcast episode in the form of a natural dialogue between two hosts. Normally you do this manually through the interface. My Python script does it automatically via the unofficialnotebooklmlibrary, which drives the browser through Playwright.

What happens under the hood: the script creates a new notebook in NotebookLM, adds the article URL as a source, requests audio generation, and waits for it to finish. When the episode is ready, it downloads the MP3 and deletes the notebook. All on a dedicated Google account, separate from my personal one.

One constraint: Google allows a maximum of 3 episodes per day. The script handles this gracefully, queuing the rest for the following day.

The finished MP3 lands on my server, and the script updates a feed.xml file in RSS format. I added that feed to Apple Podcasts as my own private podcast. The result: my articles show up in the queue next toTechStorie andKuźniar Technologicznie. Regular earbuds, regular walk, my own list.

Four models instead of one

I did something more interesting here than a plain "summarise this article." The pipeline uses the dual-researcher pattern:

• Researcher A hunts for specifics: numbers, claims, potential traps and inaccuracies.
• Researcher B builds the big picture: context, why it matters, what follows from it.
• Verifier compares both outputs and looks for gaps they both missed.
• Synthesizer assembles everything into a readable article with sections and key takeaways.

All four calls go through OpenRouter. Cost per article: a few cents.

I've had this pattern as a skill in my workflow for a while. I use it whenever I need to get up to speed on something fast — a research paper, library docs, a technical spec, a long RFC. Instead of reading cover to cover I get structure and key points in a minute. Porting it into the news pipeline was a natural fit.

A shallow "summarise this" gives shallow results. Two models looking at the same text from different angles, plus a verifier hunting for holes, give significantly deeper and more reliable summaries. The difference is especially noticeable with long or dense material.

What I got out of it

I stopped hoarding links. Now when I spot an interesting article I send it by email and don't think about it again. The system does its job in the evening. In the morning I have a podcast in the queue.

The summaries live on my private page as a knowledge base. Each one is linked to the original URL, date, and topic tags. When I'm working on a related problem two weeks later, I have somewhere to look. Podcast episodes are automatically deleted after 30 days to save server space. Summaries stay.

The whole thing runs without my attention, for under €0.12 a day.

Built in one evening

The NotebookLM piece took the longest, because the library is unofficial and you need to handle a few error scenarios. The pipeline itself and the RSS integration took about an hour.

I've had the dual-researcher pattern ready for a while — I just copied the prompts from another project. Wiring it up to OpenRouter is a few lines of code.

If you already have a mailbox, Python, and a Google account, you're halfway there.