screencapture

macOS ships screencapture, a fast CLI for grabbing the screen. Use it to see what the user sees when working on native GUI software. Web apps belong in a browser-automation tool; this skill is for everything else (SwiftUI/AppKit, Electron, Java/Swing, Qt, Tk, games, system UI, Finder, Xcode, etc.).

Core workflow

1. Capture to the scratch directory with a unique filename. Pick the directory in this order: the session scratch directory when the host names one in its instructions (Claude Code calls it the scratchpad directory), else $TMPDIR, else /tmp. A session directory is isolated from other sessions and from the user's project, so files there cannot collide or clutter shared space. Set OUT to that path and use it in every command below.
2. Read the image with the Read tool — Agent is multimodal and will see the PNG inline.
3. Delete the file immediately after reading. The user wants temp space kept clean; never leave screenshots lying around.

Filename convention: $OUT/screencap-$(date +%s).png. The timestamp prevents collisions if you capture multiple times in one task.

Picking the right mode

Goal	Command
Whole main display, no UI sound	screencapture -x -m $OUT/screencap-<ts>.png
All displays	screencapture -x $OUT/screencap-<ts>.png (one file per screen — macOS appends suffixes)
User picks a window or region interactively	screencapture -x -i $OUT/screencap-<ts>.png
Specific rectangle (x,y,width,height in points, origin top-left)	screencapture -x -R 100,100,800,600 $OUT/screencap-<ts>.png
Specific window by id	screencapture -x -l <windowid> $OUT/screencap-<ts>.png
Capture after delay (e.g. to let a menu open)	screencapture -x -T 3 $OUT/screencap-<ts>.png

Always pass -x so the shutter sound doesn't startle the user.

-i is interactive: it blocks until the user clicks/drags. Only use it when you genuinely need the user to point at something — for autonomous "show me the app's main window" checks, prefer full-screen (-m) or a known rect (-R).

To find a window id for -l, you can run osascript -e 'tell app "System Events" to get id of windows of process "<AppName>"' or use CGWindowListCopyWindowInfo via a small helper; this is rarely worth it — full-screen capture is usually fine.

Full flag reference

usage: screencapture [-icMPmwsWxSCUtoa] [files]
  -c         force screen capture to go to the clipboard
  -b         capture Touch Bar — non-interactive modes only
  -C         capture the cursor as well as the screen. only in non-interactive modes
  -d         display errors to the user graphically
  -i         capture screen interactively, by selection or window
               control key — causes screenshot to go to clipboard
               space key   — toggle between mouse selection and window selection modes
               escape key  — cancels interactive screenshot
  -m         only capture the main monitor, undefined if -i is set
  -D<display> screen capture or record from the display specified. -D 1 is main display, -D 2 secondary, etc.
  -o         in window capture mode, do not capture the shadow of the window
  -p         screen capture will use the default settings for capture. The files argument will be ignored
  -M         screen capture output will go to a new Mail message
  -P         screen capture output will open in Preview or QuickTime Player if video
  -B<bundleid> screen capture output will open in app with bundleid
  -s         only allow mouse selection mode
  -S         in window capture mode, capture the screen not the window
  -J<style>  sets the starting of interactive capture
               selection       - captures screen in selection mode
               window          - captures screen in window mode
               video           - records screen in selection mode
  -t<format> image format to create, default is png (other options include pdf, jpg, tiff and other formats)
  -T<seconds> take the picture after a delay of <seconds>, default is 5
  -w         only allow window selection mode
  -W         start interaction in window selection mode
  -x         do not play sounds
  -a         do not include windows attached to selected windows
  -r         do not add dpi meta data to image
  -l<windowid> capture this windowsid
  -R<x,y,w,h> capture screen rect
  -v         capture video recording of the screen
  -V<seconds> limits video capture to specified seconds
  -g         captures audio during a video recording using default input.
  -G<id>     captures audio during a video recording using audio id specified.
  -k         show clicks in video recording mode
  -U         Show interactive toolbar in interactive mode
  -u         present UI after screencapture is complete. files passed to command line will be ignored
  -H         capture content in HDR
  files      where to save the screen capture, 1 file per screen

Cleanup is mandatory

The user keeps temp space tidy. After every Read of the screenshot, immediately rm it in the same logical step. Don't batch — capture, read, delete, one screenshot at a time. If you need a sequence of screenshots, still delete each one as soon as you've consumed it.

Example flow:

OUT="${SCRATCH_DIR:-${TMPDIR:-/tmp}}"   # SCRATCH_DIR = the session scratch path from the host's instructions, if any
screencapture -x -m "$OUT/screencap-1715300000.png"
# → Read tool on $OUT/screencap-1715300000.png
rm "$OUT/screencap-1715300000.png"

If a capture fails (e.g. no Screen Recording TCC permission), screencapture writes a black or empty image rather than erroring loudly. If the Read shows a black/empty frame, tell the user to grant Screen Recording permission in System Settings → Privacy & Security → Screen Recording for the terminal app running Agent, then retry.

When NOT to use this skill

• Web pages, web apps, anything in a browser → use the browser automation skill.
• The user asks for a screen recording (video) → screencapture -v works but is rarely what's wanted; ask first.
• You just want to know what app is focused → osascript -e 'tell app "System Events" to name of first process whose frontmost is true' is faster and doesn't need TCC.