This short shows one real run of a pipeline that turns a single line of text into a narrated, captioned vertical video.
The brief was "Synthwave drive through a neon city", run with zero API keys (mock scene art and an offline, draft-quality voice). Here is what happens between the brief and the MP4, in plain terms:
An LLM writes the shots. The brief becomes a list of scenes, each with an image prompt, narration and timing. The output is validated strictly, so a broken JSON reply fails cleanly instead of crashing the render. 2. Providers make the pictures and the voice. Each provider is a small adapter, so you can swap a mock for a real image or voice service. 3. A job queue renders in parallel. Failed steps retry with backoff, and a crashed job can be recovered or retried with one click. 4. ffmpeg puts it together. Each scene gets a slow Ken Burns zoom, scenes stretch to fit the voice-over, and captions are burned in, timed to the narration. The output can be 16:9, 9:16 or 1:1.
If you want the ffmpeg side in detail (the 9:16 crop, the zoom filter and the caption timing), I wrote it up here: Building an AI Shorts pipeline with ffmpeg. Disclosure: the pipeline in the clip is our paid kit, the AI Video Pipeline Kit ($49, full source, Next.js + TypeScript + ffmpeg, 37 tests).