Home · Open source · Writing · Contact
11 · Rubric and data
Generated-UI design rubric
A written rubric for judging the visual design of interfaces that AI tools generate, with six frozen prompts and 24 graded outputs.
CC BY 4.0 rubric and data · MIT scripts · Frozen 27 October 2026
Read it as a page · Repository on GitHub
AI tools can now build a landing page or a dashboard from one sentence. Whether the result is good is still a judgement call. I wrote my judgement down: one holistic question, "Would you deliver this to a client as it stands?", and six dimensions, each a whole number from 1 to 5 with written anchors at 1, 3 and 5.
The rubric was frozen at version 0.2.0 on 27 October 2026. The six prompts were frozen before any tool was run. Both are published with their SHA-256 hashes so anyone can check they did not change.
The rubric
Give the holistic grade first, then the six dimensions. Thirteen written rules cover the cases that cause arguments, such as stand-in imagery, a page that runs off the side of a phone screen, and what to do when you cannot tell.
| Name | Question | |
|---|---|---|
| H | Client-ready | Would you deliver this to a client as it stands? |
| D1 | Hierarchy | Where does the eye go first, second and third, and does that order match what the page is for? |
| D2 | Layout and spacing | Are alignment, grid, spacing and density consistent, at both widths? |
| D3 | Typography | Do scale, weight, pairing, line length and line spacing make the text easy and pleasant to read? |
| D4 | Colour and contrast | Is the palette cohesive and restrained, does colour have a job, and is text legible against its background? |
| D5 | Brief fit and distinctiveness | Does it look made for this brief, or for any brief? |
| D6 | Finish | How complete and consistent is it in the small things: content, imagery, icons, components, edges? |
What is in the data
Six prompts: a coffee subscription page, a freelance invoicing dashboard, a leather boot product page, a project management pricing page, a children's book illustrator portfolio and a lighthouse restoration article.
Twenty-four outputs from five tools (Claude, ChatGPT, Lovable, v0 and Bolt), each graded on all seven rows. Beside each one are the measured facts: fonts in use, the lowest text contrast, overflow at phone width and defect flags. Each page was judged from two full-page screenshots, one at 1280 px wide and one at 375 px.
What to distrust
- The grades are an AI judge's. I reviewed 11 of the 24 and agreed with all of them. The other 13 I have not reviewed. That is sign-off, not independent grading.
- An AI agent ran the tools. It drove my own logged-in browser. Some of those tools' terms forbid automated access. I chose that and I carry the risk.
- Six of 30 slots are empty. One v0 page returned nothing, one Bolt page needs a backend, and four Bolt runs hit a free-plan limit that I did not pay to get past.
- No averages and no ranking. The set is small and the judge is one judge. A one-point difference between two items is within what the judge itself marked unsure.
What it is not
It is not a benchmark, a leaderboard or a scorer, and it says nothing about which tool is best. It is not offered as training data. It sells nothing. I run a design studio that sells design and AI-search work, and you should know that.
Cite, and licence
Rish Sadh, "Generated-UI design rubric", release v0.2.1 (rubric version 0.2.0), 2026. DOI for all versions: 10.5281/zenodo.23689021.
Rubric text, prompts, grades and notes: CC BY 4.0. Scripts: MIT.
If you grade the same prompts and disagree with a grade, I want to see where. The disagreement is the useful part.
Also on Reidify: reidify.design/open-source/generated-ui-rubric/
By Rish Sadh, founder of Reidify, an AI-first design studio in Mumbai.