Paper Collage Motion

One pipeline, two halves: analyze a reference into an editable spec, then build usable stills and a motion handoff that Codex can support. The look is halftone paper-collage / stop-motion editorial craft: flat bold color field, printed-dot elements, clean cut edges, soft layer shadows, one sharp idea per scene, and motion that assembles the scene piece by piece.

Speak in the user's language. No em dashes anywhere in user-facing output; use periods, commas, or line breaks. Keep labels and copy free of emoji unless asked.

Hard defaults

• Use $imagegen for raster image generation or edits. When a still, reference-style variation, product mockup, or generated bitmap is needed, load and follow the installed system image generation skill. Prefer its built-in tool mode unless the user explicitly asks for its CLI/API fallback.
• Do not depend on proprietary ad/video services. Do not call service-specific generators, upload workflows, remote asset registries, or overlay tools unless those exact tools are available in the active tool list and the user explicitly asks to use them.
• Generate stills before motion. The finished still is the source of truth. Motion planning is based on the approved still and the JSON spec.
• Handle exact labels locally when needed. Image generation can attempt simple burned-in labels, but exact text is unreliable. For important labels, generate the artwork without text and add the label with a deterministic local method such as SVG, HTML/CSS, canvas, Pillow, or an existing design system.
• Do not promise AI video rendering by default. If no Codex-native video generation tool is available, deliver a still set, storyboard, motion prompt pack, and folder handoff. If the user later provides rendered video files in a folder, inspect, compare, rename, package, or edit them with local tools.
• Keep prompts self-contained. Every prompt must stand alone. Include the subject, style, palette, composition, label plan, constraints, and any reference-image role each time.

PHASE A: ANALYZE

Run this whenever a reference image is in play. If the user refers to an image but none is attached or available on disk, ask for it rather than guessing.