1 · Keyframes — $0.08 each

Nano Banana 2 composes each shot from reference images: the avatar's face pack + the real product photo. Same person, exact product, every frame. The Judge inspects every keyframe before any video money is spent — wrong hands, extra people, phones in frame, wrong label = blocked and retried once, then stopped.

2 · Voice — ≈$0.02 per video (budget)

ElevenLabs speaks the script in the avatar's locked voice — the same voice in every video that persona ever makes.

3 · Talking video — $0.02/second (budget)

FlashTalk animates the keyframe to match the voice audio. Audio-driven means no prompt guessing — no invented phones, no third hand. Premium mode swaps in Happy Horse ($0.14/s) for maximum realism.

4 · B-roll — $0 (budget)

Ken Burns motion (slow zoom/pan) on product keyframes, cut over the continuous voice. Standard TikTok grammar.

5 · Assembly + final judge

ffmpeg stitches, burns the phone-shot feel in (grain + loudness), and the Judge samples frames from every clip — anything off gets flagged on the card before you post it.

Recipes — Hamza's brain

Prompt packs and judge rules live in recipes/. Editing a recipe upgrades every future video, zero code changes.