Browser Screenshot

Capture screenshots of the web application for visual verification during UI/CSS iteration.

Prerequisites

No project setup or link is needed. The project must already have Node.js, the playwright package in a node_modules reachable from its dependency directory, and a matching Chromium. Never install packages or browsers from this skill; when the helper reports them missing, hand the user its command (npm install playwright && npx playwright install chromium) to run themselves. Existing project scripts/screenshot links keep working; leave them alone.

Locate the helper

The canonical helper is ~/.agents/skills/browser-screenshot/scripts/screenshot.sh. Only when it is absent or SKILLS_DIR is set, resolve and validate the installed root read-only first. A nonempty absolute SKILLS_DIR is authoritative; never replace an invalid override silently. Otherwise prefer the canonical root, falling back to the Claude alias root only when this skill unit is absent. Legacy Codex-only roots require explicit SKILLS_DIR.

set -eu
if [[ -z "${SKILLS_DIR:-}" ]]; then
  SKILLS_DIR="$HOME/.agents/skills"
  if [[ ! -d "$SKILLS_DIR/browser-screenshot" ]]; then
    SKILLS_DIR="${CLAUDE_SKILLS_DIR:-${CLAUDE_HOME:-$HOME/.claude}/skills}"
  fi
fi
[[ "$SKILLS_DIR" = /* ]] || { echo "SKILLS_DIR must be absolute" >&2; exit 1; }
for resource in SKILL.md scripts/screenshot.sh; do
  [[ -f "$SKILLS_DIR/browser-screenshot/$resource" && -r "$SKILLS_DIR/browser-screenshot/$resource" ]] || {
    echo "Missing browser-screenshot resource: $resource" >&2; exit 1;
  }
done
[[ -x "$SKILLS_DIR/browser-screenshot/scripts/screenshot.sh" ]] || { echo "screenshot.sh is not executable" >&2; exit 1; }

Missing resources stop the run; do not repair the installation. Reuse the verified absolute path in later tool calls, since shell variables may not persist between them.

Usage

/browser-screenshot
/browser-screenshot http://localhost:3000
/browser-screenshot http://localhost:3000 mobile

Instructions

1. Determine the URL to screenshot:
   • If $ARGUMENTS contains a URL, use it
   • Otherwise use the SCREENSHOT_URL env var if set
   • Default: http://localhost:3000
2. Determine the viewport:
   • If $ARGUMENTS contains mobile, use mobile viewport (375x812)
   • If $ARGUMENTS contains tablet, use tablet viewport (768x1024)
   • If $ARGUMENTS contains a custom size like 1024x768, use those dimensions
   • Default: desktop viewport (1280x900)
   • If $ARGUMENTS contains both, take two screenshots (desktop and mobile)
   • Keep both dimensions under 2000px to avoid image limit errors
3. Run the installed helper directly as one command, passing the directory whose node_modules contains Playwright with -C (for example a nested web app). Omit -C when the current directory already resolves it; parent node_modules directories also count. Quote paths containing spaces, and never prefix the call with cd … &&.

~/.agents/skills/browser-screenshot/scripts/screenshot.sh -C web <url> <viewport>

With an override or fallback root, use its verified absolute path instead: "<verified-root>/browser-screenshot/scripts/screenshot.sh" -C "<dependency dir>" <url> <viewport>. A relative output path (third argument) stays relative to the invocation directory.

Permissions: the frontmatter grant is Claude Code metadata matching the canonical ~/.agents/... prefix, the convention other skills use; exact matching of an expanded or override path is not verified, so expect a prompt there. Pi does not enforce allowed-tools, and Codex applies its own sandbox and approval policy. Report any prompt rather than widening permissions.

Screenshots are saved to a fresh private directory (mode 0700) created per run, unless you pass an explicit output path as the third argument. The script prints the full path it wrote — read it from the output rather than assuming a location. Screenshots may capture authenticated pages, which is why the directory is not a predictable shared one.

4. Read the screenshot using the available image-reading tool to view the captured image. This is the critical step because the model can inspect the captured UI directly.
5. Analyze the screenshot and report what you see, especially:
   • Layout issues (overlapping elements, alignment problems)
   • Spacing/margin problems
   • Color or styling discrepancies
   • Responsive behavior (if both viewports captured)
6. If the user asked you to fix a specific CSS issue, compare what you see against their description and suggest or apply fixes.

Notes

• The script uses waitUntil: 'networkidle' with a 15s timeout — if the page loads async content, the screenshot will wait for it.
• Screenshots capture the viewport only (not full-page scroll) to stay within the 2000px image dimension limit.
• For rapid iteration: make the CSS change, then run /browser-screenshot to verify, repeat.