Screenshot Markup

screenshot-markup

Draw on a screenshot — circles, rectangles, arrows, captions, spotlights, blur, cursors, zoomed insets — to mark up exactly what the user asked about. Most commonly used right after a Playwright MCP screenshot, but works on any local image.

This skill is deterministic rendering only. The intelligence — interpreting the user's prompt, locating the focus region in the screenshot, and choosing the right style — lives in you, the calling model. This script just turns your decisions into pixels.

When to use this skill

Trigger when the user wants visual emphasis on part of a screenshot. Examples:

• "Highlight the Subscribe button in this screenshot."
• "Make the pricing table stand out, blur the rest."
• "Point out the error banner with an arrow."
• "Circle the timezone dropdown — it's tiny, also zoom in on it."
• "Make a tutorial-style image showing where to click next."
• After Playwright MCP returns a screenshot, when the user references a specific UI element they want to draw attention to.

Do NOT use this skill for: cropping, OCR, generic image editing, or composing multiple screenshots together.

The four styles

Pick the style that best fits the user's intent. If they didn't specify, infer from these heuristics:

You can also chain styles by running the script multiple times (e.g., spotlight then add a caption with shape).

Locating the focus region

You have two ways to tell the script where to focus:

• Area — --region "x,y,w,h" (pixel coordinates, top-left origin). Best for buttons, tables, panels, banners.
• Point — --point "x,y" plus optional --radius N. Best for "click here" or pinpointing an icon.

How to obtain coordinates:

1. From Playwright MCP: if the user identified the element by selector, ask Playwright to return the element's bounding box (element.boundingBox()), then pass it as --region. This is the most reliable path — prefer it.
2. By visual reasoning: if no bbox is available, open the screenshot, identify the element visually, and estimate its bounding box. Coordinates are in image pixels (not CSS pixels — they match the rendered PNG).
3. Ask the user if you're unsure and the screenshot is dense.

CLI reference

Script path: scripts/annotate.py (run with python3). Always pass absolute paths for --input and --output.

python3 scripts/annotate.py \
  --input  /abs/path/to/shot.png \
  --output /abs/path/to/out.png \
  --style  {spotlight|blur|shape|cursor-zoom} \
  [--region "x,y,w,h"  | --point "x,y" [--radius 80]] \
  [--caption "Text label"] \
  [--color "#E53935"]

Style-specific flags:

• spotlight:
  • --dim-opacity 0.65 — how dark the outside gets (0.0–1.0). Default 0.65.
• blur:
  • --blur-radius 8 — Gaussian radius. 6–12 is typical.
• shape:
  • --shape {rect|circle} — default rect for areas; circle is auto-used for points.
  • --stroke 4 — outline width in pixels.
  • --arrow-from {tl|tr|bl|br} — optional arrow pointing AT the shape from the given direction. The arrow originates ~140px from the shape edge in that direction (NOT from the page corner) and terminates at the shape edge with the arrowhead just outside the box.
  • --arrow-length 140 — distance in pixels from the shape edge to the arrow tail.
• cursor-zoom:
  • --zoom-factor 2.0 — magnification of the inset.
  • --zoom-corner {tl|tr|bl|br} — optional override. By default the inset is placed adjacent to the target on the side with the most free space, which keeps the visual connection tight. Pass --zoom-corner only when the surrounding UI is crowded and pinning to a page corner reads cleaner.
  • --no-cursor — skip the macOS cursor overlay (zoom only).
  • --no-zoom — skip the inset (cursor only).
  • --cursor-color {auto|dark|light} — cursor color. Default auto samples the background around the target and picks the variant with better contrast (dark cursor on light UIs, light cursor on dark UIs). Override with dark (also accepts black) or light (also accepts white).
  • --cursor-size 56 — cursor size in pixels (square). Default 56. The script auto-shrinks when the target is smaller than the cursor so the cursor doesn't swallow the UI element it's pointing at.

⚠️ --caption is opt-in

--caption writes a text label onto the image itself. Only pass it when the user explicitly asks for a text label — phrases like "label it X", "add the text Y", "write 'Z' on it", or "make a tutorial card with a caption".

For pure pointing requests like "highlight the Send button", "circle the headline", "point at the dropdown", or "draw on the toggle" — omit --caption entirely. The annotated image will be shown to the user with a description underneath it (in chat, in a viewer, in a doc); duplicating that as text on the image is visual noise.

Worked examples

Area highlight with spotlight + caption:

python3 scripts/annotate.py \
  --input  /tmp/shot.png --output /tmp/out.png \
  --style spotlight \
  --region "120,300,420,180" \
  --caption "Pricing tiers"

Click target with cursor + zoomed inset (auto cursor color):

python3 scripts/annotate.py \
  --input  /tmp/shot.png --output /tmp/out.png \
  --style cursor-zoom \
  --point "640,512" --radius 30 \
  --zoom-corner br --zoom-factor 2.5 \
  --caption "Click 'Subscribe'"

Forcing white cursor on a dark UI screenshot:

python3 scripts/annotate.py \
  --input  /tmp/dark-shot.png --output /tmp/out.png \
  --style cursor-zoom \
  --region "1015,715,90,45" \
  --cursor-color white

Zoom inset only (no cursor) — magnify a tiny element:

python3 scripts/annotate.py \
  --input  /tmp/shot.png --output /tmp/out.png \
  --style cursor-zoom \
  --region "1180,40,32,32" --zoom-factor 3.5 \
  --no-cursor

Cursor only (no inset) — "click here" pointer without magnification:

python3 scripts/annotate.py \
  --input  /tmp/shot.png --output /tmp/out.png \
  --style cursor-zoom \
  --region "1015,715,90,45" \
  --no-zoom

Tutorial-style red rectangle with arrow and caption:

python3 scripts/annotate.py \
  --input  /tmp/shot.png --output /tmp/out.png \
  --style shape \
  --region "200,200,300,80" \
  --shape rect --color "#E53935" \
  --arrow-from tr \
  --caption "Enter your email here"

Soft-focus blur on a panel:

python3 scripts/annotate.py \
  --input  /tmp/shot.png --output /tmp/out.png \
  --style blur \
  --region "60,80,520,300" \
  --blur-radius 10

Decision flow you should follow

1. Confirm the user has a screenshot and a focus target (ask if either is unclear).
2. Get the region or point coordinates — prefer Playwright boundingBox() over visual estimation.
3. Choose a style using the heuristic table above, or honor the user's explicit choice.
4. Run the script. Show the output path to the user.
5. If the user wants a different style or a tweak (e.g., different color, more dim), re-run — it's cheap and deterministic.

Notes

• The script writes PNG and accepts most common input formats (PNG/JPEG/WebP).
• All coordinates are clamped to the image; out-of-bounds regions won't crash.
• Cursor assets live in assets/. SVG files are the canonical source of truth (versioned for the repo); PNG files are pre-rendered at 4× supersampling and are what the script actually loads at runtime. Two variants are bundled: macos-cursor.png (dark cursor, used on light UIs) and macos-cursor-white.png (light cursor, used on dark UIs). The variant is auto-selected based on background brightness; if neither asset is present, the script falls back to a generated arrow cursor in the chosen color.
• Pillow ≥ 9 is required (uses rounded_rectangle, textbbox).