Claude Opus 5 Claims It Can See Bad Design

Writing CSS was never the hard part. Making the result not look sad was.

AI has been able to spit out front-end code for a while now. Give it a prompt, get a working page. The buttons click, the layout renders, the code runs. And yet the result often looks like it was designed by a committee of accountants trapped inside Microsoft Excel. Nothing is broken. Nothing is good either.

Anthropic says Claude Opus 5 changes that equation. It claims Opus 5 is significantly better at understanding screenshots, charts, diagrams, and visual interfaces, and better at recreating designs from a reference. That part is not new territory for AI. The interesting claim is what comes after the building.

Inspecting Its Own Work Is the New Skill

According to Anthropic, Opus 5 is supposed to look at what it just built, compare it against the reference image or existing interface you gave it, and correct the differences it finds. Not just generate. Inspect. Judge. Fix.

That is a different kind of task than writing markup. It means noticing that the spacing is off, that the mobile layout collapsed into nonsense, or that a button looks like it escaped from a website built in 2007. Any human designer catches those problems in about half a second. Whether a model can do the same thing consistently is the actual test here, not whether it can write a flexbox rule.

The code was never the hard part. Knowing something looks wrong is.

What Gets Tested Next

This is exactly the kind of claim that needs a real test rather than a demo reel. Hand Opus 5 an actual visual design. Have it recreate the interface. Check it across multiple screen sizes, because mobile is where most AI-generated layouts fall apart first. Then watch what happens when it reviews its own output.

Does it catch the broken spacing and the ugly button, or does it confidently declare victory over a pile of CSS that only technically works? That gap, between functioning and looking right, is where a lot of AI-built interfaces currently live.

If Opus 5 closes that gap even partially, it changes what "AI can build a website" actually means. Not just code that runs, but a result you would not be embarrassed to ship. Worth running the test yourself before you take Anthropic's word for it.