video-generated

Shots that do not exist as footage or UI. Everything spoken, written or branded stays out of the generated picture: text, captions, UI and the end card are layered on top in video-build. video-studio's laws apply to the content of every shot.

Input: the film's brief.md (../video-studio/references/brief.md). It owns copy, claims, tone, material, moments and private data; this skill owns the implementation. If the brief cannot be met, say so to video-studio - never change it silently.

Setup. An Atlas Cloud account with balance and ATLAS_API_KEY in ~/.config/marketing-agent-kit/.env - how to get one: references/atlas.md "Get a key" (and the kit's connections/atlas-cloud.md). Python 3 only, no packages. Every call costs money.

A=${CLAUDE_SKILL_DIR}/scripts/atlas.py
python3 $A upload keyframe.jpg
python3 $A --jobs <film>/jobs run video shot.json --tag s01 --out <film>/shots

references/atlas.md holds the client, the model table with billed prices, the doctrine, looping and transparent shots, and the assembly traps. On any new model: one paid call, then read the dashboard, before the second.

Camera terms for a shot's prompt (one move per shot; term + subject + direction + duration): ../video-moves/references/prompts.md.

Adapt it

• Put your own character sheets, set plates and brand ground colours in your film folder and reference them; the model follows the reference, not the prose.
• Re-check every price on your dashboard after the first call of each model, and update the table in references/atlas.md with the date.
• If you use another provider, keep the doctrine and the traps, and swap scripts/atlas.py for that provider's client with the same upload / submit / poll / download shape.
• Add your own realism rules (what your customers' world looks like) to video-studio's "Brand laws (yours)" - they bind every generated shot.