Your coding agent just became a motion-graphics director

What it does

Vox Director is an agent skill that orchestrates a full video pipeline through Atlas Cloud APIs and local ffmpeg. You give a coding agent a single topic—say, Mexican street food—and it produces a narrative beat map, generates collage-style keyframes, animates them, adds AI voice-over and music, and assembles a final mp4. Two human gates keep you in control: you approve the story beats and pick the visual style from a generated bake-off, while the agent handles everything else.

The interesting bit

The pipeline treats each narrative beat as a finished collage poster first—complete with torn paper, halftone dots, and bold headlines—and only then brings it to life with motion. That separation of look from motion is what preserves the Vox aesthetic; if the keyframe is flat, no downstream video model can rescue it. The skill also handles three distinct input modes: pure generation from a topic (B-roll), restyling existing talking-head footage frame-for-frame (A-roll), or anchoring a still photo as a photographic sticker (C-roll).

Key highlights

• Runs end to end on Atlas Cloud models (nano-banana-2 for keyframes, gemini-omni-flash for motion, xai/tts-v1 for narration, minimax/music-2.6 for BGM) plus local ffmpeg assembly
• Two mandatory human approval gates: the narrative beat map and the visual style bake-off
• Supports A-roll and C-roll workflows that preserve real human faces or product photos without redrawing them
• Fetches live model IDs from the Atlas Cloud API before running, so the skill adapts to provider drift automatically
• Ships with prompt guides, fourteen narrative arc templates, and voice rosters in references/

Caveats

• Requires a paid Atlas Cloud API key, a compatible coding agent, and local ffmpeg/ffprobe with Python—this is not a standalone web app
• The advanced element-level motion engine is optional; most projects use the default “living poster” animation path
• Exact model behavior and availability depend on Atlas Cloud’s live API, which the skill queries at runtime

Verdict

Developers and content teams already living inside agent workflows will find this genuinely useful for rapid, stylized explainers. If you want a point-and-click video editor or would rather not manage API keys and agent orchestration, this is not your tool.