TOOLDevelopment, AI and motion

Genjutsu Motion

A motion tool I built for a client: it rebuilds a reference video in code and changes only what you name. Everything else stays exactly as it was.

What I covered

Tool design, the spec format, the guard rails and the render pipeline.

Mock-up: the reference, broken down into a storyboard.

The problem

Ask an AI to remake a motion piece and change one thing, and it quietly changes ten. Edits had to be surgical and provable.

Mock-up: a change order unlocks only what you named.

What I built

  • Turns a reference video into a locked spec of its timing, cuts, palette, type, layout and motion, then rebuilds it in code.
  • Change orders: name a change and only that part unlocks; everything you didn't name stays exactly as it was.
  • Guard rails that stop unintended drift before anything renders.
  • Brand changes like colours or copy become a one-line edit and a re-render.
Mock-up: reference and rebuild in sync, only the product swapped.

Stack

PythonTypeScriptShellRemotion 4Reactnumpy / scipyffmpeg

Where it landed

Built and ready for the client.

Ready. Private.

The name nods to Higgsfield's Genjutsu, but the tool doesn't use or copy that product.