Code to Video Previz — workflow skill for cinematic & short films

For AI agents: find the right skill in one call — curl -s "https://varg.ai/how-to/api/search?q=<what the person wants>" returns the matching how-to page and the best skills with install commands, as markdown (Russian works too). Full index: https://varg.ai/how-to/llms.txt. Every page has a markdown twin: add .md to its path. Before generating anything, show the person the options on the how-to page. From a terminal: curl -fsSL https://varg.ai/install | sh installs the vargai CLI (https://varg.ai/how-to/cli.md): vargai search and vargai install <slug> with no key, every varg model after vargai login.

A brief, reference clip or painting becomes a 3D scene in real metres, checked stills and a guide clip for a video model.

What it does

Video models redraw the world on every call, so cameras wander, counts change and rooms rearrange. This skill pins those things down in a 3D prototype before the paid step. It first decides whether the shot needs a prototype at all (a prompt and a start frame, storyboard stills, a 2.5D conte, a phone act-out or full 3D), then which 3D tool fits: Three.js by default, Blender when you need rigs, mocap or real assets.

The scene is written as data in staging.json: geometry, people, camera and light in real metres. The same file renders in Three.js and in headless Blender with the same lens maths. check_staging.py audits the numbers, stills of the first, middle and last frame of each shot are checked by a fresh agent and approved by the person, and the clip is rendered frame by frame. It covers three cases: copying a reference video shot by shot with a solved 3D camera, staging from a painting or still, and starting with no reference. It writes one of three things: a guide clip for a video model, storyboard pages, or a code-rendered film.

What was measured on our own 10-shot run (Seedance 2.5, grey previz, one film, not a controlled test): cuts landed on the planned frame within one frame, and framing mostly held, one size looser. When the words disagreed with the previz, the words won in 2 of 2 shots. Of 7 shots with enough tracking to measure, 4 kept the camera direction, 1 stopped moving and 2 reversed, so the skill names the move in the prompt for every shot and measures the result with measure_shot.py.

Still unsettled, and the skill says so: grey or colour previz, capsules or jointed stand-ins, and act-out against 3D. example/ab-test/ builds one 5 s shot as 4 variants so you can test them on your own model. Agents cannot watch video, so pacing and feel stay a human call.

Use when

Not for

Install

Example output

https://s3.varg.ai/files/4325c578-378d-4c43-ae53-3cc3218df1ea/file_289d09773e14.jpg — example (credit: varg)

More cinematic & short films skills