Turns a photo or a written scene into a handcrafted 3D layered-paper collage — torn edges, visible paper grain, soft cast shadows — while keeping the subject recognizable. Use when you want a warm, tactile craft look instead of another digital-painting filter.
Most style filters make a photo unrecognizable. This one is built around not doing that.

The skill turns a photo, or a written scene, into a handcrafted 3D layered-paper collage: torn and deckled edges, visible paper grain, cast shadows between layers, a restrained scatter of paper flowers and stars. What makes it worth cataloguing is not the look — plenty of prompts get you paper texture — but the preservation checklist it carries: pose, gaze, expression, hairstyle, clothing silhouette, subject count and composition are all named as things the model must not touch. It changes the medium and leaves the person alone.
A single SKILL.md — no scripts, no reference library, no data. Its value is entirely in how it writes the prompt: an eight-field skeleton (use case, primary request, subject, composition, style/medium, scene/backdrop, lighting/mood, constraints), a preservation list for edits, and an unusually specific exclusion list. That last one is the part that actually holds the style together: it names glossy plastic, clay, fabric, wood and metal as textures to avoid, because those are exactly what a model drifts toward when you ask for “handmade”. It also bans all text, letters, numbers and watermarks outright.
Written as a Codex skill — it installs to the Codex skills directory and calls the built-in image tool. Nothing in the method is tied to that, though: the skeleton and both lists transfer to any image model you like.
It will replace your background with white paper, and that surprises people. The skill says to use a clean white paper backdrop unless you ask for something else. Applied to an edit, that means the park your dog was standing in simply goes away:

This is the skill working as designed, not failing — but if you want the setting kept, say so in the brief. It is the one instruction most worth overriding.
The style holds; the likeness wobbles. Across our runs the craft layer never broke once — paper grain, torn edges, visible stacking, soft shadows, and not a single stray letter or watermark anywhere. The preservation layer is softer. Re-running the same brief moved a beagle’s ears shorter and widened its face until the breed read as a cocker spaniel; another run turned a blurred lakeshore into defined mountain peaks and pulled the mouth into a firmer expression. For anything where the point is that it looks like that particular dog or person, generate a few and pick.
Aspect ratio is not yours to choose. Edits roughly inherit the source photo’s shape, but a text-only generation came back square no matter what the prompt asked for. Budget a crop.


For finding a prompt structure to start from rather than a style to apply, see AI Image Prompts.
Not ours — install from the source repository. Copy the skill folder into your Codex skills directory. Note that the repository carries no licence file, so treat it as all rights reserved: read it, run it, but do not redistribute it. It costs nothing itself; whatever your image provider charges is the whole bill.