There is a gap in how most people use AI to build websites. The AI writes the code, saves the file, and tells you it is done. But it never actually looked at the page. It has no idea whether the button works.
It is a bit like someone cooking a meal from a recipe and never tasting it. Everything was followed correctly, and it might still be inedible.
What this skill does
webapp-testing lets Claude open a real web browser and use your site the way a person would.
It can click things, type into boxes, fill in forms, take screenshots, and read the hidden error messages a browser records when something breaks.
It does this using a tool called Playwright, which is software that controls a browser automatically. You do not need to learn it. The skill handles that part.
Why it changes things
Without it, fixing a bug goes like this. Claude changes the code. You reload the page. You look at it. You describe what is wrong. Claude changes the code again.
You are the eyes. Every single round trip needs you.
With it, Claude loads the page itself, clicks the thing itself, and sees the error itself. It can go round that loop several times before it needs you. You come back when there is something worth an opinion.
The difference is biggest for problems you cannot spot by reading code. A pop-up hiding behind another box. A button that does nothing because it was set up before it existed. A layout that only breaks on a narrow phone screen.
Things worth asking it to check
Vague requests get vague answers. Specific ones work well.
“Open this page at phone width and tell me if anything sticks out sideways.” Content spilling off the edge on a phone is one of the most common website faults, and one of the easiest to miss on a laptop.
“Submit the contact form with the email box empty and tell me what a visitor sees.” Error messages get written and then never actually looked at.
“Load the homepage and list every error the browser reports.” Takes thirty seconds. Most sites fail it.
“Press Tab repeatedly and tell me if you ever lose track of where you are.” That is how people who cannot use a mouse move around a page, and it is very often broken.
Ask for numbers, not opinions
This is the single most useful habit.
“Does this look right?” is a weak question. Looking right is a matter of taste, and you will get a vague answer.
“How wide is this box, and is the page wider than the screen?” is a strong question. It has an actual answer. The page is either too wide or it is not, and there is nothing to interpret.
Whenever you can turn a question into a measurement, do it.
What it is not
This is for checking a site you are building on your own computer. It is not a full testing system for a big team, and it does not run automatically every time someone changes something.
What it replaces is the check you were going to do yourself and probably skip. Does this actually work, in a real browser, at this screen size.
Pair it with frontend-design for the building, and you get a loop that ends with something you have seen working, rather than something that merely saved without complaining.
All of Anthropic’s skills are free to read at github.com/anthropics/skills.
Join the discussion