What problem does it solve? When users upload images as design feedback, agents often misread them: copying red arrows and callout bubbles into the product UI, or treating every image as a rough sketch. This Skill classifies the artifact first, then separates genuine design evidence from annotation-tool styling so only the user's actual intent reaches the implementation. ## Core Features & Use Cases - Artifact classification: Distinguishes annotated screenshots, clean screenshots, low-fidelity sketches, design references, and unknown images using accompanying words, visible layers, filename provenance, and metadata. - Signal extraction: Pulls complaint, target area, direction, content, and topology from an image while discarding annotation styling such as arrow colors, callout bubbles, and screenshot-tool chrome. - Interpretation readback: Produces a structured readback of the classification and extracted intent before the first edit, and routes stated match intent to visual-inspection or open direction to taste-test. - Use Case: A user uploads a CleanShot capture with a red arrow pointing at a cramped header. The Skill reads it as an annotated screenshot, keeps the complaint and target, ignores the arrow's color and styling, and adds spacing using the product's existing design language. ## Quick Start Upload an annotated screenshot or sketch and ask the agent to interpret it as design feedback before making any UI changes.
