matching-a-reference

Visual work iterated against a picture rather than an adjective. The skill is how an agent reaches the reference workflow: start it with the work and the asker's own words for its look, then claim and answer its requests. The run makes reference images with the image_gen service, pictures the work beside each, and goes round closing the differences until none is left.

The failure it prevents: an agent asked for something "stunning" that judges its own frames against the word, sees they fall short, and has nothing that says what to change.