onetake's real contribution is measuring continuity, not just demanding it

The central claim here is architectural: most AI-generated motion videos fail not because individual shots look bad, but because nothing connects one shot to the next. Scenes replace each other. onetake's answer is to treat every boundary between beats as the primary design problem — something on screen must visibly survive the transition and become the next thing. A prompt bar opens into an app window; a bar collapses into a line that shoots across a desk and becomes the first grid line of the next sequence. The camera never cuts; it anticipates, whips ahead, floats, and holds still when stillness is what makes the preceding move land.

What makes this more than a stylistic preference is that continuity is measured. A verification script called probe.py records what appears in every frame, identifies each boundary, and asks what carried across it. Films where beats simply replace each other fail automatically. The README shows the progression: an earlier version of the launch film scored 0.00 on the carry metric and was rejected; a second pass reached 0.40, still rejected; the third pass hit 0.75 and passed. The final onetake launch film scored 0.83. The same oracle also rejects uniform shot cadence, absent stillness, clipped audio, fast moves without motion blur, and subjects that leave the frame.

The production approach is deliberately anti-shortcut. Product interfaces are rebuilt in HTML from screenshots rather than screen-recorded, so the camera can push into them at any zoom without pixelation. Motion is drawn from a library of roughly 38 named moves — springs, carries, collisions, camera behaviors — each defined as a pure function of time with a measured speed curve, not a hand-rolled ease. Motion blur is computed by averaging multiple captures across an open 180-degree shutter in linear light, so fast moves smear rather than strobe. Sound is placed from the film's own events into a single reverb space, with music ducking under hits. Output is deterministic: any seek returns the same frame, with drafts at 1080p30 and finals at 4K60.

Ten case films ship with the repo, each with a breakdown covering concept, beat sheet, numbers, and what was rejected and why. Source code for the films themselves is not included. The skill installs into Claude Code or other agent environments that read a skills directory, and the dependency list is substantial: Playwright with Chromium, NumPy, SciPy, Pillow, Matplotlib, optionally OpenCV for reference analysis, FFmpeg, and Node. Narration adds faster-whisper and Kokoro TTS.

The license is PolyForm Noncommercial 1.0.0 — free for personal, educational, and research use, commercial use explicitly prohibited. That constraint matters for anyone evaluating this as infrastructure for client work or product marketing, which is precisely the use case the README pitches.