Daily Digest

I save too many articles and read too few of them. Pocket, Instapaper, Shiori. The app doesn’t matter. The backlog always wins.

So I automated the part I was skipping: actually going through it. Every morning, a pipeline turns yesterday’s saved articles into a 5–10 minute spoken briefing I can listen to over coffee. My reading list, distilled.

The pipeline runs automatically overnight and retries on failure, so it’s ready when I wake up.

1. Fetch — Pull the last 24h of saved bookmarks from shiori.sh. Everything is connected—the same reading list I manage with the Shiori iOS app feeds directly into this pipeline.
2. Script — Sonnet writes a conversational briefing that connects ideas across articles, highlights tensions, and surfaces what’s worth digging into.
3. Speak — Text-to-speech generates the audio. I’ve tested every option: ElevenLabs has the best voice quality but costs add up. OpenAI TTS is solid for the price. The best output is actually NotebookLM—two AI hosts debating the content feels surprisingly natural—but there’s no API yet.
4. Publish — Generate artwork, build the RSS feed, push to my phone. Ready to play.

The tech is straightforward. What’s interesting is the format and the choices behind it.

I didn’t want a flat summary. I wanted an opinionated debate: two hosts with deep knowledge of my interests, my tools, my process, discussing the most interesting topic from yesterday’s reading. The French bias, probably. I want friction and perspective, not a bullet-point recap.

Getting there is prompt design. The LLM needs enough context about who I am and what I care about to produce something that feels crafted for me. The pipeline is plumbing. The editorial direction baked into the prompts is where I actually spent time.

The pipeline also connects the dots across my side projects: shiori.sh as the input layer, Distill’s synthesis philosophy as the design principle, and a personal automation stack that runs itself. Each project feeds the next.