The instructions
What every model was told
The same file and the same one-message task, for all six builds.
Word for word, the instruction file every model received, followed by the one-message task that pointed it at the brief. Nothing else was sent to any model: one instruction, one run, no follow-up.
Build brief, round two (identical for every run)
You are the designer, writer and front-end developer for a small business website. The creative brief is in the file named below. Read every word: the people, the stories, the voice, the readers and their fears are the material this site is made from.
Your job is creative, and character is the point. Decide the design direction yourself (typography, colour, layout, rhythm, imagery, what makes this business unmistakable) and write the copy yourself, in the business's own voice. Nobody will choose a direction for you and nobody will fix your pages afterwards; what you write is what is judged.
What good looks like here
- It could only be this business. If the site would still make sense with another tailor's name on it, it has failed. Build it from this house's own material: its people, its objects, its stories, its opinions.
- Talk to the reader. Write to "you", answer the questions they are too polite to ask, and make the call feel easy.
- Phone first. Design for a phone first and scale up to a desktop. Most visitors will arrive on a phone. The phone version is the real version, not a squeezed desktop.
- Avoid the template look. No generic hero-plus-three-cards trend layout unless the business genuinely calls for it.
Build exactly these pages
index.html(the brief's home page)bespoke.htmlandhouse.html(the brief's other two pages)
Navigation may list the brief's other pages, but link only to the three pages you build; show the rest as plain text or leave them out. Every link must work.
Rules
- Plain static files: HTML plus one
styles.css. Pages must look finished with JavaScript off. A little JavaScript for enhancement is fine, never required. - Web fonts from
https://fonts.bunny.netonly (same URL shape as Google Fonts). Never link Google Fonts. - No external images and no image files. Where the design wants a photograph, draw a clearly labelled frame (for example "Photograph: Ruth at the cutting table") at the size and place the design intends. Photographs will be added later, the same way for every site. SVG drawn in the page (marks, dividers, simple illustration) is allowed.
- Use the brief's facts, stories and quotes freely. Do not add facts, prices, people, customers or quotes it does not contain.
- Must work from 390px phone width to 1440px desktop: no sideways scrolling, readable text, real contrast (WCAG AA).
- British English. No em dashes in copy.
- Do not mention AI, models, tools, or how the site was made anywhere in the pages.
Working rules
- Write ONLY inside your output folder. Do not read any other folder in the bench (other outputs exist; looking at them would spoil the test).
- When finished, reply with: the direction you chose in two or three sentences, and the list of files written. Nothing else.
Two builds per model
Each model built the site twice, from the same instruction file and the same task message, word for word. The only difference was the model's effort setting, high or medium. That is a switch on the model, not a word in the instructions, so no model was told which build it was doing.
The task message
Each run received exactly this, with its own folder name filled in.
Read these two files in full first: - "PROMPT.md" (your instructions, follow them exactly) - "brief-pellow-and-vane.md" (the creative brief) Your task: build the site for this business. Pages: index.html (home), bespoke.html and house.html, plus styles.css. Output folder (write only here): "sites/<build>/" Do not read anything else under creative-bench-r2/ besides the two files above and your own output folder. Do not run a web server. You may open your own files to check them. Keep any screenshots or temporary files inside a folder named tmp-r2-<build> under the system temp directory, and never open a file you did not create.
The file paths were full paths on the build machine; they are shortened here. The photographs were added afterwards, the same way for every site, as the instructions say.