Atelier: describe it, get it, keep editing it

Small businesses, creators and agencies need new video, product shots, decks and documents every week, but a separate editor for each tool is expensive, and a coding agent that writes and debugs scripts burns tokens, gives different results each run and runs code nobody checked. Atelier splits the work in two. The model decides what to make. A fixed engine decides how. You describe the piece in plain words, for example a 15 second vertical coffee shop reel with upbeat music and captions. Atelier plans it, writes the script, makes the voice, music and pictures through Vultr Serverless Inference, and renders it while you watch each step. The model never writes code and never touches file bytes. It writes a short JSON spec of about 200 tokens. A validator rejects anything outside the allowed operations, and the engine renders the spec the same way every time. Results are cached, so a repeat render costs no tokens. The first result is not a dead end. It opens in an editor with a timeline, trim, filters, stickers, drawn masks, zoom, animated looks, clip order and captions, with a live preview. Every hand edit and every AI edit goes through the same engine and is saved as a new version. The same service also edits images, audio and Office and PDF documents. It runs on one Vultr Cloud Compute VM under Docker Compose, with every model call going through Vultr Serverless Inference and every file mirrored to Vultr Object Storage. Model output is never run as code, and shell commands, where they run at all, run in a throwaway container. The live demo and the code are public.