Design Director

Design Director is a Claude Code skill I run whenever an agent builds something a person will look at. It follows Anshu Chimala's Discover, Define, Deliver process with one structural rule: the agent that writes the code never grades it. A separate critic reads only a screenshot and a one-sentence aesthetic statement, three times with a fresh context each time, and the median score decides what the next round fixes. New directions start from a random string made by a shell script, because a model asked to be unique keeps returning the same page. The work comes in three sizes: a spot check that costs a few cents after any screen change, a review when a surface slips, and a full redesign for new surfaces. Nothing past the spot check runs until I say yes.

The rules came from measurement, not taste. On a bench of eleven real sites, Sonnet picked the better of two versions twenty times out of twenty, yet its absolute scores on the same image swung from 1 to 7, so it runs the cheap comparison and never gives a score. Folding a checklist into the score dropped every real site to a 1 or a 2, so the checklist and the score are reported apart. Handed an inspiration photograph, the critic ignored it in 31 of 33 critiques. A written brief, produced once by a stronger model, gave every critic the same reading on every run.

The plates show the skill designing its own repository card. Three seeded directions scored 6, 5 and 6, and I chose the eye chart, the risky one. Three critic rounds then failed to beat 6. With no brief to anchor them, the critics reversed themselves between rounds: use the wide frame, then delete the side column; make the claim readable, then shrink it past reading. The card that shipped is round three after a subtraction pass. The record of that plateau is published with the skill, because a design tool that only shows its wins is making the mistake it was built to catch.