Nehad Al Ghadri on stage, holding a microphone

The Rabbit Hole

A personal app I built that turns one topic a week into three narrated episodes of about 10 minutes, released Monday, Wednesday, and Friday. Each one comes with chapter timestamps, show notes, a podcast feed, and an automatic archive to my notes. A script on my Mac writes the episodes, and a Next.js web app plays them.

Episodes per weekly topic
3
Per episode, ±30s tolerance
~10m
Pushes per week
6

Next.js 16, Vercel Blob, Claude CLI, Kokoro TTS, ffmpeg, launchd, ntfy

I can find text on almost any topic, but rarely narrated long-form. I wanted to go deep on one topic a week, one I pick, whatever the week demands, as radio-quality audio I can listen to in the car or on a walk. I also wanted the professionalism of a real podcast (measured chapters, show notes, a valid RSS feed), and I wanted the whole thing to run itself once a week with two taps. The Rabbit Hole is that system.

What I built

A four-layer architecture I now use as the reference pattern for every phone-first personal app I build.

Worker (Mac)
A launchd poll plus a Sunday picker kicks the week off. A bash orchestrator runs two-pass Claude content generation for each of three days, passes every draft through a ruthless editor quality gate (avg ≥ 8.0, min ≥ 7), renders each section through Kokoro TTS, and stitches the episode with ffmpeg earcons and per-episode artwork.
Queue (Vercel Blob)
Pending topic, current state, episodes, chapters, artwork, feed. The worker writes and the webapp reads, never the other way round.
Webapp (Next.js 16)
A /pick route to choose the week’s topic from the phone. A /library route to browse every past week. A /library/<week>/day/<n> page with an audio desk, chapter navigation, and day-to-day nav. Passphrase middleware, no OAuth, single user.
Notifications (ntfy)
One topic, six pushes a week, no per-section progress spam. Sunday trigger, worker starting, week ready, then Mon/Wed/Fri daily release reminders.

Measured chapters

Word-count-estimated chapter timestamps drift 10 to 60 seconds over a 10-minute clip, which makes them unusable for a scrubbable player. I inject ###CHAPTER### markers in the input text, render each section through Kokoro separately, sum cumulative audio samples, and record startTime = cumulative_samples / sample_rate per marker. The player reads that array. Tapping a chapter lands on the exact word, not an estimate.

The quality gate

A two-pass ruthless editor scores every draft on six axes. A draft below the bar gets one tightening regeneration pass with the critique injected. Most drafts pass first or second try. It’s the same lesson I’ve learned on every AI pipeline I’ve built: a smaller model with a strict critic outperforms a larger model with no critic.

What it delivers

Each week I get in-depth audio on one topic, and I do listen to it. The first real run, Week 17, was three episodes on Naturalistic Decision Making, generated and shipped end-to-end with zero API cost (everything runs through the Claude CLI under my Max subscription). Seven minutes of compute produced thirty minutes of listening I’d otherwise never have found time to make.

All work