You spot a broken layout in your app's preview. The signup button sits half off the card on mobile, and you want Cursor to fix it. The slow way is to screenshot, save the file, drag it into a note, type out a description, then explain to Cursor where to look. The fast way takes a browser tab and about a minute. Here is the whole thing, step by step.
Step 1: Open a capture in the tab you already have
Go to Start a new review. Nothing to install, no extension, no signup. Click Capture screen and the browser asks which window or screen to share. Pick the tab or window showing your app. Cobalt Capture draws the current frame to a canvas.
Outcome: you have a still image of the exact screen state you want Cursor to fix, without leaving the browser.
Step 2: Crop to the thing you mean
Drag a rectangle around the button and its card. Cropping is the only image edit here, and that is on purpose: a tight crop tells the agent where to look better than an arrow drawn on a full-page screenshot. If the whole frame matters, skip the crop and keep it all.
Outcome: one item now holds a cropped still of the broken button. Cut everything Cursor does not need to see, following the same logic in what to crop in and out for an agent.
Step 3: Say what's wrong instead of typing it
Click into the comment field and dictate. In Chrome or Edge the browser's speech recognition transcribes as you talk, so you can say something like: "On mobile, the primary signup button overflows the right edge of the card. Constrain its width to the card and add sixteen pixels of horizontal padding." Talk in the words you use out loud and it reads cleaner than most people type.
If you are in Firefox, dictation is not available there, so type the same thing. Either way, be specific about the element and the fix. Vague comments waste a turn; concrete ones get applied. Dictating is faster than typing for most people, and there is more on that in dictating review comments instead of typing them.
Step 4: Drop a pin on the exact spot
If the crop still leaves any doubt about which control you mean, add a numbered pin pointing straight at the button. The pin becomes part of the item and shows up as a coordinate reference in the output. When a screen has three buttons that look alike, this is what stops Cursor from editing the wrong one. See pinning the exact spot a reviewer means for when it earns its place.
Outcome: the item now names the element, the problem, and the fix, with a visual anchor to match.
Step 5: Publish and grab the markdown URL
Click Publish. The review saves and gets a short public URL like /r/your-slug. That link is readable by anyone you send it to, and it is the human-facing version. But the one Cursor wants sits at /r/your-slug/markdown, plain-text markdown with your cropped still, your dictated comment, and any pins laid out in order.
You published without an account. An anonymous review is kept for 30 days, and if you sign in later with Google or a magic link, it becomes yours permanently. No account is required to get the URL you need right now.
Step 6: Paste the URL into Cursor
Switch to Cursor and paste the markdown URL into the chat, followed by a short instruction: "Apply the fix described here." Cursor reads the markdown, sees the cropped image, the element description, and the exact change you asked for, then edits the right file. Because the input is structured text rather than a video or a loose screenshot, the agent does not have to guess.
The canonical, kept-current version of this handoff lives in the guide on getting feedback into Cursor. If you are wiring this into a repeatable routine, that is the page to bookmark. For a broader map of the same idea across other agents, the Cursor page and the AI coding agents hub cover the variations.
Why this beats a raw screenshot paste
You can paste an image straight into Cursor, and sometimes that is enough. But an image alone gives the agent no filenames, no element names, and no stated fix. It has to infer all of that. A markdown URL carries the picture and the words together, which is why a video does not work here either: a Loom clip gives Cursor nothing to read. That contrast is spelled out in why Loom doesn't work with agents.
The same published link doubles as a record. If you want to check that a teammate opened it, the review's analytics page shows visit counts. If Cursor's change looks right, you can mark the item resolved and move on. One capture, one link, two audiences: a person who reads the public page, and an agent that reads the markdown.
Try it on the next bug you spot. Open a tab, capture the screen, say the fix out loud, publish, and paste the markdown URL into your Cursor chat. Time it. It should land under a minute.