From a sentence to a shipped ad, in one conversation.
Under the chat sits a production pipeline that was tuned on real renders. Here is exactly what happens, in order, and why.
Understand the product
The Director fetches your page: text, images, the buttons you use, your colours, your logo. For apps and services it screenshots the real interface. A vision model turns that into a grounded product model — what it is, who buys it, three claims that are actually true. If you have no brand kit, it creates one with the logo and palette.
Decide the ad
Platform is inferred from your words (“for Instagram” means 9:16, 8 seconds). The CTA comes from your own site. Then one planning pass writes the concept, three hooks in the styles that perform this year — relatable problem, bold claim, curiosity, demo-first, and social proof only when your page has real proof — plus the voiceover and a one-line reason for every choice.
Write the shot for the engine
Video models can't spell, ignore timing markers and render “transforms into” as a smear. The Director translates the plan into a single continuous shot in the language engines execute well: one subject, one action, one camera move, positive phrasing only. On-screen text is lifted out to be composited later, sharp and spelled correctly.
Design the frames
An opening frame and a CTA end card are generated, cropped to the exact aspect, scored 1–10 by a vision reviewer, and the real logo file is stamped onto the card. You see both with their scores before the video renders.
Render in parallel
The content take and the end card render at the same time; music starts composing the moment you press go. Clips are stitched locally with a crossfade at 30fps, the voiceover and music are mixed to the exact length, captions and the logo mark are burned in, and the mix is mastered to the -14 LUFS the feeds expect.
Review, then edit by chat
Four frames of the finished ad are scored for product fidelity and artefacts. The verdict is posted to the chat. “Fix it”, “different hook” or “shorter” re-renders only the affected scene, with the cost quoted first.
Measured, not promised
- 21 seconds to read a site and propose a priced plan
- 82 seconds to design and score both frames
- About 4 minutes from “go” to the finished 8s ad
- Quote equals charge: 120 credits shown, 120 charged