Annotate a screenshot for ChatGPT or Claude

"The button in the top-right… no, the other one" is not a prompt. Mark the regions on the screenshot itself, paste the generated text into the chat next to the image, and the AI knows exactly where everything is — in the screenshot's own pixels.

How do I annotate a screenshot for an AI chat?

1. Take the screenshot, then paste it straight into the editor with Ctrl+V — no saving to disk first, no upload.
2. Draw over the parts you're going to refer to. Rectangle over a button or panel, an arrow pointing at what should move where, a point marker on a small target, freeform around an irregular problem area. Label each region ("save button", "cluttered toolbar") and add a note where context helps ("overlaps the sidebar on narrow screens").
3. Hit Done and copy the prompt-ready text tab.
4. In ChatGPT or Claude: attach the same screenshot, paste the text under it, and ask your question — "Redesign the area marked 'cluttered toolbar' without touching 'sidebar nav'."

What does the pasted text look like?

Labeled regions for image "app-screenshot.png" (1440x900 pixels).
Coordinate system: origin at the top-left corner;
x increases rightward, y increases downward.

## save button (1 region)
- rectangle at x=1296, y=18, width=104, height=40
  note: primary action, currently hard to find

## sidebar nav (1 region)
- rectangle at x=0, y=64, width=240, height=836

## broken spacing (1 region)
- arrow from (612,388) to (704,352)

Plain language plus numbers — every LLM reads it without special handling. When code or an agent is on the other end instead of a chat window, use the JSON tab: the same regions in exact pixels and normalized 0–1 coordinates, versioned and documented field by field in the docs.

Why not just describe where things are in words?

Because relative descriptions are ambiguous — "the icon next to the other icon" has three plausible readings on a busy toolbar, and the model picks one silently. Going the other direction is worse: vision models are unreliable at producing pixel coordinates themselves, so asking "what are the coordinates of the save button?" gets you plausible-looking numbers that can be far off.

Human-marked regions solve both. You point once, precisely, and the AI reasons about locations it can trust. It's the same principle behind our image coordinates for AI workflow, applied to the screenshot conversations you're already having.

Can the AI mark the screenshot for me?

Yes. Click the purple Auto button and AI suggests labeled regions — on a UI screenshot it looks for nav, buttons, and content areas. Type an optional hint first ("just the form fields") to steer it, then adjust labels and boxes before you export. Auto is opt-in: in normal use the screenshot never leaves your browser, and when you do click it, the image is sent solely for that analysis and never stored. It's part of Premium — one $5/month plan with multi-image projects and batch export.