Human design-taste signal for OpenAI — rating generated web layouts and motion against a trained eye, and writing down why one reads better than another so the judgment can be learned rather than guessed. Across Codex and GPT-6 Astra.
A model can generate a thousand layouts. It can’t yet feel which one is right.
That gap is the job. I sat on the human side of it for OpenAI — judging generated web design and motion the way a designer does, instantly and then explaining it, so the preference behind the instinct could become training signal instead of a shrug. Not "this one", but why this one: the hierarchy lands, the spacing breathes, the motion has intent.
Web design evaluation
Pairwise scoring of generated pages — hierarchy, spacing, type, and whether the layout actually did the job it was made for.
Motion & taste
Judging motion design and overall taste — the parts of a piece that are felt before they are reasoned, and the hardest to put into a rule.
Written rationale
Every rating backed by a short reason, so the model learns the thinking, not just the verdict.
The signal fed real systems.
The evaluations contributed to OpenAI models including Codex and GPT-6 Astra — a small, human reference point for subjective visual quality inside systems that otherwise reason their way to everything.
