Claude can see – screenshots and the browser

"The button is sort of shifted, and on mobile the table wraps." Try putting that into words precisely enough for someone without a screen to find the bug.

You don't have to. Claude Code can read images and even drive a browser itself.

🖼️ Pasting screenshots

Take a screenshot, press Ctrl+V in the terminal (Cmd+V also works in iTerm2), add what's wrong. Claude spots layout bugs, reads error messages from dialogs and can build a first component from a sketch or a design mock-up.

🌐 Claude in Chrome

With the Chrome extension and claude --chrome, Claude controls your browser:

• open pages, click, fill in forms
• read console errors and the DOM state
• test a complete flow, e.g. from sign-up to the confirmation page
• record the whole thing as a GIF, perfect for the pull request

This also works in apps you're logged into. At logins and CAPTCHAs, Claude pauses and hands control back to you.

💡 My favourite workflow

"Fix the bug in the shopping cart and then check in the browser that it's gone." Claude changes the code, reloads the page, reads the console and only then reports "done". The loop closes without me.

A word of caution:

The browser runs with your sessions and permissions. If you let Claude work in a logged-in admin area, you should know exactly what you're asking it to do. Claude in Chrome also requires a direct Anthropic plan; it doesn't work with an API key alone.

My rule of thumb:

• Hard to describe? Screenshot instead of a paragraph.
• Changed a web frontend? Let Claude check it in the browser itself.
• Production admin, payments, customer data? Only with you sitting next to it.