ambitionOS Studio

Long-form video is where automation usually falls apart: motion design, pacing, captions, loudness, and packaging each demand judgment, and stitching them by hand caps output at whatever one editor can survive.

The interesting question was whether two AI systems could split the work the way a studio does — one decoding what good motion looks like, the other engineering a pipeline that reproduces it deterministically.

The two-model workflow is on disk, not a slide. Gemini analyzed a reference edit and decoded its motion system into a 27-preset library (PiP, frame, and overhost presets with rotation and rhythm rules), then authored two pipeline PRDs. Claude Code triaged those PRDs in an adopt/reject table and built the result: a 6,034-line beat-composition engine governed by four JSON Schema contracts.

The engine is deliberately LLM-free at runtime — it must reproduce a frozen golden baseline (the shipped 332-second cut, 23 cues) within 35 milliseconds on every merge. Scripts come from a local gemma4:26b, so the unattended cron line produces complete private drafts with zero cloud-LLM spend. HyperFrames renders the karaoke captions and overlay plates from HTML.

Every episode passes automated QC (loudness, A/V drift, black-frame and caption-coverage checks) and then a human approval gate on a local review server. Uploads are browser-automated with per-field screenshot verification — title, description, thumbnail, tags, and the AI-content disclosure confirmed from pixels, not return codes.