Product and innovation · Explainer
Looks Right, Feels Wrong
AI tools can now build a working screen in an afternoon. In our view, building the screen has become cheap, and judging whether people can use the screen has become the scarce skill. This page gives you five questions for judging any screen a team or vendor builds with AI. The five questions group twenty well-known design principles, so you can ask the questions without learning twenty names. At the end you check one screen from your own work and get a list of questions for the people who built the screen.
Guess first
What does the applause tell you?
Tap an answer
A vendor builds a new claims screen for your call centre in two days with an AI tool. At the demo, six people say the screen looks great. What does that praise tell you about how easy the screen will be for a call-centre agent to use?
Five questions
Check the claims screen
This is the vendor’s screen. Tap each question to see where the screen fails that question, read the principles behind the question, and apply the fix. The screen changes with each fix.
The orange dashed outlines mark what the open question points at.
The twenty principle names come from the video credited in the sources. The grouping into five questions, the claims screen and every fix are our own illustration.
Read the principles with care
Twenty principles are not twenty laws
Only some of the twenty come from measured experiments, such as Fitts’s Law and Hick’s Law. Others describe how the eye groups shapes, and others are rules of thumb that designers share. Treat all twenty as good defaults, and accept a break from one when the team can say why.
The principles can pull against each other. The Von Restorff Effect says to make one thing look different, and the Law of Similarity says to make things that do the same job look alike. Choosing which principle wins on a given screen is the judgment this page is about.
Numbers need care too. Miller’s 1956 paper put the limit of what people hold in mind at about seven items, and later work puts the limit nearer four (Cowan, 2001). The lesson is to group items, not to cap every menu at seven.
The twenty leave out accessibility, which is whether people with low vision, a screen reader or no mouse can use the screen. The Web Content Accessibility Guidelines (W3C, 2023) cover accessibility, and in Ontario the Accessibility for Ontarians with Disabilities Act requires larger organizations to meet an earlier version of those guidelines on their websites. Ask about accessibility as a sixth question.
The Attention Bottleneck page shows why the person reviewing AI work becomes the limit, and What Surrounds an AI Agent shows what else to check before a company assistant goes live.
Your turn
Check one screen
Pick one screen from your own work, such as a form a vendor built for your claims team or a dashboard a colleague made with an AI tool. For each question, choose whether the screen passes. Choose “Not sure” whenever you cannot say, and the page will turn each “Not sure” and “No” into a question for the people who built the screen. Your answers stay in this browser and are never sent to us.
For the team that builds
Give the AI tool the rules first
If your team builds screens with an AI tool such as Claude or ChatGPT, the team can paste these rules into the tool’s standing instructions, so every new screen starts closer to usable. The rules below are our own starter set, written from the five questions.
Once the screen is built, ask for a check in a separate request: “Check this screen against the eight rules. List the problems from worst to least bad, and do not change the screen yet.” Then fix the worst problems first. We suggest a separate request so the AI tool reviews the finished screen instead of grading its own plan.
Use it this week
Ask one question before the applause
At the next demo of a screen built with AI, ask one question before anyone comments on the look: what should a first-time user notice first? Then ask someone who has never seen the screen to finish the main task while you watch, and count the moments the person hesitates.
Building a screen is cheap now. Judging the screen is the skill, and five questions cover most of the judging.
Judging AI work is one part of digital intelligence, and the 30-day course builds the rest of that skill one day at a time.
Sources
- A short video by Adam Hayes (@adam_ha_yes) on Instagram, “UX Laws your Claude Needs”, 19 October 2026. The names of the twenty principles come from this video, as summarized for us in writing. The video offers files of instructions to viewers who comment, and we have not seen those files.
- Jon Yablonski, Laws of UX (O’Reilly, 2020; second edition 2024) and lawsofux.com, for the plain definitions of the principles.
- Masaaki Kurosu and Kaori Kashimura, “Apparent usability vs. inherent usability”, CHI 1995, on attractive cash-machine screens being rated easier to use.
- Walter Doherty and Arvind Thadani, IBM, 1982, on response times under about 0.4 seconds.
- George Miller, 1956, and Nelson Cowan, 2001, on how many items people hold in mind.
- Daniel Kahneman and colleagues, 1993, on the peak-end rule.
- Jon Postel, 1980, on accepting input in many forms, first written for computer networks.
- W3C, Web Content Accessibility Guidelines 2.2, 2023.
- The five questions, the claims screen, the fixes, the worksheet questions and the eight starter rules are our own.