Reel Pipeline Series
6 weeks. 6 posts. One pipeline that turns a Substack article into a 9:16 reel — narrated, captioned, animated, on a $0 GPU.
Join a waiting list to get 10 USD discount
Part 1 — The whole machine in one diagram. Architecture overview. What goes in (a published article), what comes out (a 9:16 MP4), and the five stages between. The vocabulary you’ll need for everything that follows.
Part 2 — Teaching Haiku to write a screenplay. How I turn an article into a 3–5 scene JSON script with Claude Haiku 4.5. The prompt, the JSON contract, and the cache layer mean a re-render costs zero AI credits.
Part 3 — Why I shipped the free TTS I don't use. Kokoro as the default (80MB, runs on CPU, $0, no signup), ElevenLabs one dropdown away. The SceneAudio shape that made the swap a one-afternoon job — and why a paywall before the first reel is the real bug.
Part 4 — Captions that read along with my voice. ASS karaoke format, libass, and the math of pinning subtitles to a fixed mobile anchor so they never jump between scenes. Word-by-word red fill as I read.
Part 5 — Ken Burns is the cheapest motion you can ship. Three SVG layers per scene rendered through Sharp at 2160×3840, animated with FFmpeg’s zoompan, and fused under the captions. The motion baseline before AI got involved.
Part 6 — $0 AI reels on a 12GB GPU. The finale. I swapped the Ken Burns slide deck for real per-scene i2v animation. Wan 2.2 14B plus a Lightning LoRA, running on a home PC over Tailscale. Zero marginal cost per reel. (coming)
