The first result wasn't where I stopped
I asked Claude Opus 4.7, ChatGPT 5.5 and Gemini 3.1 to build a website selling gummy vitamins. I gave them the same starting prompt and opened the results one by one.
I liked Claude's first attempt more than the others. With ChatGPT and Gemini, I was looking at things I thought felt generic. ChatGPT's version hadn't followed my request for two pages properly, and it didn't give me the one-bottle purchase option I expected.
Then I asked all three to use a frontend skill and improve what they had built. The second round was more interesting, especially Gemini's result.
What I looked at first
The first screen matters because that's what a new visitor sees when they open the page. But I wasn't only looking at whether it was pretty.
Could I get to the product? Could I choose what to buy? Had the model actually made the product detail page I asked for?
I opened the pages and talked through what I liked and what I thought was missing. My reactions are opinions from one test, not evidence that one design would sell more than another.
What the skill added
In the recording, I describe a skill as a file with a name, a description of what it does, examples and guidelines. It gives the coding assistant more specific direction about how to do a job.
I used Anthropic's frontend design skill as the reference and asked the other assistants to adapt it to their own setup. I also asked them to keep the same local links so I could come back and compare the results.
You can see the second instruction from around 5:00.
That second pass changed more than one thing: I added the skill and gave the models another chance to revise. So I wouldn't claim this isolates the skill's effect by itself.
Looking at the second versions
Claude's revised page looked better to me, although I still wanted to change parts of it. ChatGPT's version added details I noticed while looking at the purchase area. Gemini surprised me enough that I said I would buy from that design.
That was a reaction to the page, not a sale. I didn't run a conversion test or demonstrate a completed store launch.
For my next attempt, I'd keep a short checklist beside the prompt: the pages I need, the products and options that must exist, the information that must be accurate, and the purchase steps I want to test. Then I'd use the design instructions to improve the presentation without losing those requirements.
Watch the first and second versions to see what changed and what I still didn't like.
Resources
- Anthropic's frontend design skill: the design guidance discussed in the video. The current file may differ from the version used during recording.
- Claude, ChatGPT and Gemini: the model families named in this recording.