Cursor already let you point at part of your running app and tell its agent what to change instead of describing the change in a sentence. Cursor updated that feature, called Design Mode, today, adding voice narration and a faster, more precise way of telling the agent what you mean.

What Design Mode actually does

Design Mode runs inside the Cursor browser, an in-app preview window where your running app renders while you work. From there, you can click an element on the page, and instead of writing out where it is and what it looks like, Cursor captures the element’s exact identity: its position in the code, the specific component it belongs to, and its computed styles (the actual visual styling applied to it once all your design rules are combined), along with a screenshot of that part of the screen.

That combination is what lets the agent make a targeted edit instead of guessing which button or section you meant from a text description alone.

Design Mode also supports two other ways to point at what you want changed:

  • Multi-select: reference two or more elements at once, useful when the change depends on a relationship between them, like asking the agent to make one component match another or clean up repeated content across several.
  • Drawing: circle or box a region on a frozen screenshot of the page. The annotation sits over the exact page state you were looking at, so the agent edits the area you marked, not whatever the page looks like by the time it responds.

What’s new in this update

Three things changed in today’s release. Voice input is new, so you can narrate an instruction instead of typing it while pointing or drawing. Element targeting is more precise and faster than the earlier version. And you can now send a second edit before the first one finishes, described as making it easier to multitask across several changes at once. As each queued edit completes, the app hot-reloads (it refreshes on its own to show the change, so you don’t manually reload the page) and you see it land in the running product.

Cursor recommends its Composer 2.5 model for this workflow specifically because it’s fast and strong at interface work, which matters when you’re sending edits back to back rather than waiting on each one individually.

What it requires

Design Mode needs the Cursor browser open, and you reach it through Cursor’s Agents Window. Cursor’s announcement doesn’t mention any plan restriction or added cost for the feature, and doesn’t publish one, so don’t assume it’s limited to a specific tier without checking Cursor’s own pricing page first.

The part to watch: multitasked edits still need a review

Design Mode makes it easy to fire off several UI changes in a row without waiting to see if the first one landed correctly. That’s the point of the update, but it also means you can queue a second and third edit before you’ve confirmed the first one was scoped to the right element.

The risk isn’t hypothetical: if the page re-renders between when you pointed at an element and when the agent finishes editing it (a common occurrence if the section you’re editing is dynamic, like a list that updates), the agent could act on a slightly different version of that element than the one you saw. Hot reload will show you something changed, but it won’t tell you on its own whether the agent touched only the file and component you intended.

Treat the hot-reloaded preview as a first look, not confirmation. Open the diff Cursor generates for each edit (a side-by-side view showing exactly which lines of code changed), the same review step you’d use for any agent-written code change, and check that it modified only the component you selected before you queue the next one.

What builders should do next

Try Design Mode on a small UI change you’d normally describe in a sentence, like adjusting spacing on a card component. Do it two ways: first, select the element, describe the change, and wait for it to complete before moving on. Then try it again on a different part of the page, but this time select a second, unrelated element and queue a follow-up edit before the first one finishes.

Compare the two runs on two things: correctness, whether each edit landed on the exact element you meant in both cases, and corrections, how many edits you had to manually fix or redo afterward. If queuing multiple edits produces more corrections than doing them one at a time, that’s a sign to slow down on more complex or crowded parts of a page, even though the feature lets you move fast.

Where this fits

Design Mode’s real shift is capturing exact page state, an element’s code identity plus a screenshot, instead of relying on you to describe a UI location in words. That’s part of a broader pattern in coding agents: accepting messier, more human ways of giving instructions (pointing, drawing, talking) and translating them into precise code changes on their own. The tradeoff is that you’re now trusting the agent’s interpretation of a visual signal, not an unambiguous sentence, so the review step matters as much as it ever did.

End of article