Tomorrow is someone's first day. They've asked for less motion. I won't be there to explain this.

One small instruction. A file. A folder. A confirmation.

For this version, I need the motion to stop. I still need the instruction.

I could pause it, collect screenshots, or edit the keyframes myself. A screenshot keeps an image; I want editable vector states in one Lottie file.

So I bring my original and brief to PulsePatch.

On AWS, it renders actual frames. Claude sees the brief and images, then proposes states. OpenCV measures changes and each layer's visible contribution. In this recorded run, its feedback reached a later model decision.

Then I review what stays. The document before. The move. The confirmation. I approve this sequence.

The model proposes; it doesn't write the final animation. After approval, deterministic export holds those states without continuous object movement. Vector layers remain editable. Eight checks pass on the saved file.

In the small, AI-assessed synthetic study, adaptive had six correct factual outcomes: four exports and two author questions. One-pass had four. The registered advantage rule failed. Human understanding wasn't measured.

I won't be there to explain it. But I can choose what I leave.

Created by Damian with AI coding assistants.

PulsePatch proposes. The author decides.