I tried letting an AI agent run my design crits

The premise: hand a week of design critiques to an AI agent and see what it does to the room.

I wired it into our review board, gave it the same rubric the team uses, and let it speak first in every session. The goal was not to replace the crit — it was to find out whether a machine could name the thing everyone in the room was already feeling but nobody had said out loud yet.

A wall of printed UI screens pinned in a grid in a dark studio, one lit by a spotlight — The crit wall, week one — every screen the agent saw before we did

What actually happened

It was ruthless about consistency and blind to intent. It caught spacing drift and contrast failures faster than any of us, then confidently praised a flow that quietly broke for anyone who had used the product before. The lesson was less about the agent and more about us: the parts of a crit worth protecting are the ones that require having lived with the problem.

A monitor in a dark room showing a design file covered in dense machine-generated annotations — One pass, 214 annotations — spacing drift it caught in seconds

Where the grid breaks

By Friday the pattern was obvious. The agent grades against the grid; the room grades against the person on the other side of the screen. Both matter — but only one of them can be automated, and it is not the one that decides whether the product feels right.

A strict off-white grid on black, one cell slipping outside the grid — Consistency is a grid. Intent is the cell that slips out.