Visual Reference Generation Agent Skill

Visual reference generation

Arguing about an image costs an hour. Arguing about a build costs a sprint. Generate the picture

first.

Before generating

Settle these, or the output is decoration:

• Surface and platform — web page, native mobile screen, dashboard. These are not the same

problem at different aspect ratios: touch targets, native chrome, and scroll behavior change what

a good layout is.

• What it optimizes for — one conversion, one task completion, one first impression. Stated, so

the image can be judged against something.

• Content reality — real headline lengths, real data volumes, real edge cases. A concept built

on three-word labels collapses on contact with actual copy.

Generating

• One concept per image. Tiling several ideas onto one canvas makes them impossible to compare

or iterate separately.

• Generate genuinely different directions, not variations of one. Three near-identical options

is one option presented three times.

• Include the states that will exist: a populated view and an empty one, at minimum.

Web versus mobile

Web — the fold is a real constraint but not a hard one; horizontal space allows genuine layout

choices; hover exists. Design for a range of widths, and decide what the narrow case does.

Mobile — thumb reach dictates where primary actions sit; native navigation patterns are

expectations, not suggestions; there is no hover, so affordance must be visible. Design the scroll,

not the screenshot.

After generating

Say explicitly what in the reference is direction and what is placeholder. A developer

handed a concept will otherwise implement the lorem ipsum faithfully.