Twelve steps from idea to upload.

Llama 3.3 70B (Replicate)

We feed your channel theme into an LLM tuned for hook-first YouTube ideas. It returns titles that sound like they earned their click.

Perplexity sonar-pro

Before the script gets written, we pull a real research dossier — quotes, dates, deal sizes, sources. No LLM-hallucinated 'fun facts.'

Llama 3.3 70B

An 8+ minute long-form script with a strong cold-open, 12+ scene markers, and a paid-off ending. Tagged for shotting and voiceover chunking.

Llama 3.3 70B

Each scene gets an image prompt, a duration, and the voiceover line that plays over it. This is the manifest the rest of the pipeline runs against.

gpt-image-2 (fal.ai)

Twelve+ cinematic key frames at 16:9, each anchored to its scene. We retry with sanitized prompts if a frame refuses.

Seedance 2.0 image-to-video (fal.ai)

Each storyboard frame is animated into a ~5 second motion shot. Camera moves, parallax, subject motion — all driven from the image.

Veo 3.1 first-last-frame-to-video (fal.ai)

Between every pair of clips we generate a tied transition that takes the last frame of one shot and the first frame of the next. No hard cuts.

ElevenLabs v3 (fal.ai)

The script is voiced in the channel's chosen voice. Per-scene chunks so we can sync timing with the storyboard.

ffmpeg on Fly.io

Clips, transitions, and voiceover are stitched into a single MP4 with loudness normalization and a clean master audio bus.

gpt-image-2

Long-form thumbnail in 16:9 and a portrait variant for shorts — each with hook text designed to earn the click, not describe the video.

Llama 3.3 + ffmpeg

The script's most replayable beats are identified, cut from the rendered assets in portrait, and captioned. 3–5 shorts per long-form, every time.

YouTube Data API

Long-form and shorts go up on your channel. A scheduled job rotates thumbnails every 48h for the first two weeks and crowns the winner by CTR.