Output shape
The skill does not merely say that a screen is cluttered or confusing. It names the user path, identifies the cognitive load in each step, prioritises the highest-harm issues and proposes the smallest change that would make the path easier.
That makes the audit useful to both sides of a product conversation: designers get a clearer diagnosis, and engineers get changes that can actually be implemented.
The problem
Ask a language model to review an interface and you get plausible, generic, unfalsifiable advice: improve the hierarchy, add whitespace, make the call to action stand out. None of it is wrong and none of it is grounded in why users behave the way they do.
The underlying research does exist — Hick's law on choice time, Miller on working memory limits, Fitts on target acquisition, Jakob's law on transferred expectations — but it is not what a general-purpose model reaches for by default.
The approach
Package the framework rather than the conclusions. The skill supplies the lenses — perception, attention, memory, cognitive load, decision-making, error — and the analysis is conducted through them, so a recommendation carries a stated mechanism and can be argued with.
Cover the whole surface of interface work, not just landing pages: forms, onboarding, navigation, settings, dashboards, checkout. These are where cognitive load actually accumulates and where most real usability cost sits.
MIT licensed and public, because a design framework that nobody can inspect is just another opinion.
What was hard
Cognitive psychology is easy to cargo-cult. "Hick's law says fewer options" is a real finding flattened into a slogan — the law describes reaction time over equally probable choices, which is not the same as an argument for hiding functionality. Encoding the reasoning rather than the slogan was most of the work.
The other risk is producing an audit that lists forty violations of thirty laws, which is technically thorough and practically useless. Prioritisation had to be part of the framework, not left to the reader.
Dogfooding
This site was built using it, which is why the catalog leads with clearly labeled paths rather than sixteen undifferentiated cards, why the contact form asks for the signal first instead of presenting a fourteen-field discovery questionnaire, and why live demos sit before the quote request.
It is also the cheapest possible proof of the AI UX capability listed on the home page: the artifact is public, free, and you are currently looking at its output.