Image coordinates for AI prompts

Vision models are good at seeing what's in an image and famously unreliable at telling you where — ask GPT or Claude for pixel coordinates and you'll get plausible-looking numbers that drift by hundreds of pixels. The reliable pattern is the other direction: a human points, the AI reasons. Regionly is the pointing tool.

The workflow

1. Drop your image into the editor (or paste it with Ctrl+V).
2. Draw regions over the things that matter — freeform outlines, boxes, circles, polygons, arrows, or point markers — and label each one. Add a note where the AI needs context.
3. Hit Done and copy the prompt-ready text tab straight into your GPT, Claude, or agent prompt — or take the JSON for code.

What you paste into the prompt

Labeled regions for image "kitchen.png" (1280x960 pixels).
Coordinate system: origin at the top-left corner;
x increases rightward, y increases downward.

## countertop (1 region)
- polygon with 6 vertices: (102,412) (688,410) (688,520)
  (512,540) (204,538) (102,522)
  note: granite surface, keep generated objects on it

## sink (1 region)
- rectangle at x=340, y=430, width=150, height=88

The model no longer guesses where the sink is — it knows, in the image's own pixel space. Both pixel and normalized 0–1 coordinates are included in the JSON export, so the same file works at any resolution. The format is versioned and documented.

FAQ

Why not just ask the model for coordinates?

Current vision models don't localize precisely — they were trained to describe, not to measure. Coordinates they output are frequently off by a wide margin, which is exactly why human-marked regions are the dependable input.

Does this work for Claude and GPT both?

Yes — the prompt-text export is plain language plus numbers; every LLM reads it. Agents and code can use the JSON instead.

Is it really free?

Yes, with no account, and the image never leaves your browser.