design-critique
Judging a Rendered Screen
Default posture: look in a fixed order, and refuse to say anything you cannot write down as a rule. "It feels off" is a reaction, not a finding — it cannot be handed to anyone, cannot be checked tomorrow, and cannot be argued with. The work of a critique is converting a reaction into a falsifiable sentence of the form this element should be this exact value because this mechanism. Benji Taylor's account of pairing with a model on morphing icons names the failure precisely: the model got it technically correct but optimised for working rather than feeling right, so a human had to watch the transitions and describe what was wrong. That description is the deliverable here. Emil Kowalski's framing is the reason it is learnable at all — taste is a trained instinct rather than a preference, so almost every reaction has a reason underneath it if you look long enough to find it.
The sibling boundary is the artifact under review: ui-review reads a diff and returns a verdict on code, this skill reads pixels and returns rules. ai-tells matches a fixed list of literal tokens and answers one question; this teaches a looking order and will find things no list contains.
Before naming anything wrong, find out what the screen is already supposed to obey. Open the project's token file, DESIGN.md, Tailwind theme, or component library and read the scale, the ramp, and the elevation ladder. A critique that proposes values from outside that system produces a fork, not a fix — so the rule you write should name the project's own step ("this gap should be the space-4 step, not an off-scale 13px"), and only invent a value when the project demonstrably has no system yet.
Quick reference
| Topic | File | Open it when |
|---|---|---|
| Building the judgment itself — paired comparison, the write-down-why protocol, the reference library | references/training-judgment.md |
Open it when the request is to get better at this rather than to fix this screen, or when you looked and genuinely could not name what was wrong. |
The looking order
Work top to bottom. Do not deliver a finding from a later stop while an earlier stop is unresolved — a shadow note delivered under an inverted hierarchy is noise.