SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsscreenshotting-apps

screenshotting-apps

Testing & quality

View repository

Updated 10-15 19:54

About this skill

Annotated Screenshot Documentation

Generate a markdown document with annotated screenshots of any web app or live website.

Prerequisites

  • agent-browser CLI installed globally (npm i -g agent-browser && agent-browser install)

Phase 1: Determine the Target

A) Live website

  1. Normalize URL (add https:// if missing): 
    agent-browser open https://example.com && agent-browser wait --load networkidle
  2. Dismiss cookie/consent banners using semantic locators (most reliable): 
    agent-browser find text "Accept" click
    Or snapshot to find the accept button ref, then click the ref directly: 
    agent-browser snapshot -i -s "[role=dialog], [class*=cookie], [class*=consent]" -c
agent-browser click @e1  # Use the ref from snapshot
    Verify it's gone: agent-browser is visible "[role=dialog]" or agent-browser wait "[role=dialog]" --state hidden
  3. For login walls or anti-bot issues, see references/live-sites.md.

B) Local dev server

  1. Find or start the dev server (check ports 3000/5173/4321/8080, or start via package.json)
  2. For self-signed certs: agent-browser --ignore-https-errors open https://localhost:<port>

Phase 2: Map the Site

  1. Open homepage: agent-browser open <url> && agent-browser wait --load networkidle
  2. Take a quick auto-annotated screenshot to survey the page — this labels all interactive elements with numbered refs: 
    agent-browser screenshot --annotate docs/screenshots/_survey.png
    Read the output legend to understand what's on the page (buttons, links, inputs mapped to @e refs).
  3. Snapshot navigation to discover pages: 
    agent-browser snapshot -i -s "header, nav, [role=navigation]" -c
    For deep DOM: agent-browser snapshot -i -c -d 3
  4. For SPAs with client-side routing: also check route definitions in the source code (grep -r "path:" src/router or similar) since not all routes may be linked from the nav.
  5. Ask the user to confirm the page list. For live sites, suggest 5-10 pages max.

Phase 3: Screenshot Each Page

Workflow per page: validate selectors → annotate → check for FAILED → verify screenshot → fix or continue.

Setup (run once before all screenshots):

agent-browser set viewport 1440 900
export AGENT_BROWSER_SCREENSHOT_DIR=docs/screenshots

Block ads and tracking scripts that create overlays or slow page loads (live sites):

agent-browser network route "**/*.doubleclick.net/**" --abort
agent-browser network route "**/googlesyndication.com/**" --abort
agent-browser network route "**/analytics.js" --abort
agent-browser network route "**/gtm.js" --abort

Resolve the annotation script path (checks common install locations):

ANNOT_JS="$(find . .claude .agents ~/skills -path '*/app-screenshots/references/annotate.js' 2>/dev/null | head -1)"
[ -z "$ANNOT_JS" ] && echo "ERROR: annotate.js not found — check skill installation"

For each page:

Step 1: Navigate

agent-browser open <url> && agent-browser wait --load networkidle

For live sites, add agent-browser wait 2000 for lazy-loaded content. Chain when you don't need intermediate output: agent-browser open <url> && agent-browser wait --load networkidle && agent-browser wait 2000

Step 2: Dismiss cookie/consent banners (live sites)

agent-browser find text "Accept" click
agent-browser wait "[class*=overlay]" --state hidden

For complex or iframed banners, see references/live-sites.md.

Step 3: Validate selectors BEFORE annotating

Always test selectors first. Failed selectors produce invisible annotations.

Use eval --stdin with heredoc for complex JS (avoids shell quoting issues):

agent-browser eval --stdin <<'EVALEOF'
const sels = ['selector1', 'selector2'];
sels.map(s => {
  const el = document.querySelector(s);
  if (!el) return 'MISSING: ' + s;
  const r = el.getBoundingClientRect();
  return 'OK: ' + s + ' (' + Math.round(r.width) + 'x' + Math.round(r.height) + ')';
}).join('\n')
EVALEOF

Quick match count check: agent-browser get count "<selector>"

If MISSING, use agent-browser snapshot -i -c to discover correct selectors and refs. Do not annotate with broken selectors. See references/selectors.md for tips.

For elements below the fold, scroll them into view first: agent-browser scrollintoview "<selector>"

Step 4: Inject annotations

Use eval --stdin to inject the annotation script safely:

agent-browser eval --stdin <<EVALEOF
$(cat "$ANNOT_JS")
annotateMulti([
  {sel: 'selector', label: 'Label', type: 'box'},
  {sel: 'selector', label: 'Label', type: 'click'}
])
EVALEOF

CRITICAL: If output starts with FAILED:, fix selectors before screenshotting. Clear and retry:

agent-browser eval "clearAnnotations()"

Annotation types:

  • click — filled circle + arrow + numbered label. Use for buttons, links, inputs.
  • box — dashed border + numbered label. Use for sections, cards, containers.
  • circle (default) — capped-radius dashed circle + arrow + numbered label. General callouts.

Colors auto-rotate: red, blue, green, amber, purple. Each annotation is auto-numbered (1, 2, 3...) for cross-referencing in markdown.

MAX 3 annotations per screenshot. Split into multiple screenshots if needed (e.g., 03a-top.png, 03b-actions.png).

Type selection rules:

  • click for small elements (buttons, links, icons) — capped circle won't explode
  • box for medium regions (cards, form groups, sidebars) — follows element shape
  • Never use circle or box on full-width elements (navbars, headers) — annotate specific items within them instead

What to annotate per page type:

  • Homepage: hero CTA, search input, one nav example
  • Product listing: one product card, filter toggle, sort dropdown
  • Product detail: price, add-to-cart, size/variant selector
  • Search results: search input, one result card, filter sidebar
  • Dashboard: one metric card, one chart, primary action
  • Table/data view: column header, filter control, data area
  • Settings/form: key form fields, submit button

Step 5: Screenshot

agent-browser screenshot <NN>-<page-name>.png

Zero-padded numbering: 01-homepage.png, 02-category.png, etc. Use --full for full-page screenshots when content extends below the fold.

Step 6: Verify screenshot quality

Read the screenshot file you just captured and check:

  • Annotations are visible and not overlapping each other
  • Labels are not cut off at viewport edges
  • The page content loaded correctly (no spinners, broken images)
  • Cookie banners or overlays are not blocking the view

If any issues: agent-browser eval "clearAnnotations()", fix the problem, re-annotate, and re-screenshot. Do not proceed with a bad screenshot.

Interactive screenshots

For search, menus, modals, hover states — use the snapshot → refs → interact pattern:

# 1. Snapshot to get refs
agent-browser snapshot -i
# 2. Interact using refs (more reliable than CSS selectors)
agent-browser click @e5          # Click search input by ref
agent-browser type @e5 "shoes"
agent-browser wait 2000
# 3. Annotate the result
agent-browser eval --stdin <<EVALEOF
$(cat "$ANNOT_JS")
annotateMulti([{sel: '[class*=suggest]', label: 'Suggestions', type: 'box'}])
EVALEOF
# 4. Screenshot
agent-browser screenshot 05-search-results.png
# 5. Verify — read the screenshot and check quality

Note: Refs inside iframes work directly — no frame switching needed for interaction.

Recording interactive flows (optional)

For complex interactions (checkout flows, multi-step wizards), record a video alongside screenshots:

agent-browser record start docs/screenshots/checkout-flow.webm
# ... perform the interaction steps, taking screenshots along the way ...
agent-browser record stop

Link the video in the markdown: [Watch checkout flow](screenshots/checkout-flow.webm)

Mobile / responsive screenshots

To document mobile layouts:

# Use device emulation (sets viewport + user agent together)
agent-browser set device "iPhone 14"

Then follow the same annotate → screenshot → verify workflow. Name files with a -mobile suffix: 01-homepage-mobile.png.

To return to desktop: agent-browser set viewport 1440 900

Note: On narrow viewports, prefer max 2 annotations per screenshot to avoid crowding.

Phase 4: Generate Markdown

Filename: docs/<domain>-screenshots.md (live sites) or docs/app-screenshots.md (local).

Use numbered references (1), (2), (3) in descriptions to cross-reference the annotation badges visible in each screenshot.

# <Site/App Name> - Visual Guide

Brief description of the site/app.

**URL:** <base-url>
**Date:** <today's date>

**Annotation key:** Red = primary actions, Blue = secondary elements, Green = content areas

---

## <Page Name>

1-3 sentences describing the page. Reference annotations by number:
"The **Search bar** (1) supports autocomplete. Use the **Category nav** (2) to browse departments.
The **Hero banner** (3) rotates seasonal promotions."

![<Page Name>](screenshots/<filename>.png)

---

Phase 5: Compare environments (optional)

When documenting differences between staging and production, or before/after a redesign:

# Visual pixel diff between two URLs
agent-browser diff url https://staging.example.com https://prod.example.com

# Diff against a baseline screenshot
agent-browser diff screenshot --baseline docs/screenshots/01-homepage-before.png

# Accessibility tree diff (detect structural changes)
agent-browser diff snapshot

Include diff screenshots in the markdown to highlight what changed.

Phase 6: Cleanup

# Remove network intercepts
agent-browser network unroute
agent-browser close

Reference