Reverse-Engineer Reference Images into Prompts

The most common mistake when writing image analysis prompts is vaguely asking, "Make something in the vibe of this photo." The AI will simply latch onto one or two dominant colors and fill the rest with generic defaults. As a result, you might get similar colors but an entirely different composition or lighting. Without clarifying why you liked the image, accurate reproduction is impossible.

The Role defines the analytical tone. Framing the AI as an art director reverse-engineering the production process yields precise observations like "hard light from top-right casting long shadows to the left" instead of subjective fluff like "warm and cozy."

The Task narrows the scope from mere appreciation to methodical deconstruction. The six pillars in the Steps paragraph—subject, composition, color, lighting, texture, and medium—form the backbone. Most people only notice two or three aspects they like until forced to review all six. Requiring visual evidence prevents hallucinations, while allowing an "Unverifiable" exit ensures accuracy.

The Format specifies two output prompts: one faithful to the original aesthetic, and one applying your modifications. Seeing both side by side helps you understand where the prompt diverged and what to adjust next. The Constraints protect copyright and privacy by extracting transferable stylistic elements without cloning specific identifiable features or proprietary names.

A generic request like this produces broad keywords such as "cozy cafe interior, warm tones, cinematic lighting." The AI grabs the most prominent colors while discarding camera angles and lighting direction. The nuanced textures and specific shadows you liked get lost, leaving you with no clear way to troubleshoot why the result looks off.