verify-android-screen

name

verify-android-screen

description

Use when verifying an Android UI state requires a real screenshot — WebView content, animations, visual fidelity (color, font, image content, alignment), or finding an element by appearance when you don't know its resource id. For everything else, prefer `verify-android-layout` first — JSON is cheaper and more precise. Triggers whenever you'd otherwise read a PNG from `android screen capture` in the main thread.

Verify Android UI via Screenshot

When to reach for this skill

Use a screenshot only when the JSON layout tree can't answer the question:

• WebView content — web markup doesn't appear in android layout
• Animations or transitions — layout may fail or return partial state mid-frame
• Visual fidelity — colors, fonts, image content, spacing, alignment polish
• Finding by appearance — locating an element when you don't know its resourceId or text

For everything else, use verify-android-layout first. A JSON dump is strictly cheaper than a vision-token screenshot for "did the element appear?" / "is the input focused?" / "did the text update?" style questions.

Why delegate

A single screenshot is a large image payload. Reading it in the main thread burns tokens fast — across an iteration loop, inline reads pollute the conversation and balloon context. Always delegate to a sub-agent. It reads the image, returns a short text answer, and the bytes never enter the main thread.

Workflow

1. Capture to a tmp path:

   android screen capture -o /tmp/<descriptive-name>.png
   

   Add --device <serial> if multiple devices are connected.

2. Spawn a sub-agent (general-purpose or Explore) with model: "sonnet" and a self-contained prompt that includes:

   • The exact file path to read
   • Specific, concrete validation criteria — what should be on screen, what shouldn't, where to look
   • The expected return format ("YES/NO + one sentence", "under 40 words")

3. Act on the text answer. Do NOT Read the screenshot yourself.

Annotated screenshots — finding elements by appearance

When you need to interact with an element that doesn't show up in android layout (or you can't identify it by resourceId):

android screen capture --annotate -o /tmp/annotated.png

This overlays numbered labels and bounding boxes on every UI element. Have the sub-agent identify the label number for the element you want, then resolve to coordinates:

adb shell input tap $(android screen resolve --screen /tmp/annotated.png --string "tap #34")

The chained form lets you tap a numbered annotation in a single command.

Example sub-agent prompts

Read /tmp/reader-after-hold.png. Verify: (a) a single large word is centered in the upper third with a red ORP letter, (b) the bottom inline-context strip is visible, (c) no code block is shown — we expect a paused image break with caption "Pipeline diagram". Answer in under 40 words: did all three pass? If not, which failed and what's actually visible?

Read /tmp/annotated.png. Find the "Sign in" button — return only its label number (e.g. #7). If multiple candidates, pick the most prominent. One token answer.

Read /tmp/webview-state.png. The page should show a logged-in user header with avatar in the top-right and a "Welcome back" greeting. Under 30 words: present or not, and what's actually shown?

Why Sonnet, not Opus

The task is narrow: read one image, check 2–3 criteria, return a sentence. Sonnet is multimodal and much cheaper than Opus for this. Haiku also works for very simple criteria. Always pass model: "sonnet" when spawning the verification sub-agent — never let it default to Opus.

Common mistakes

Mistake	Fix
Reaching for a screenshot when JSON would answer	Try verify-android-layout first; screenshots are the fallback
Reading the screenshot inline "just to check quickly"	The bytes are massive even for a quick peek. Always delegate.
Vague criteria ("does it look right?")	Spell out what should/shouldn't be on screen and where
No return-format cap	Agents return long descriptions by default. Specify "under N words" or "YES/NO + one sentence"
Letting the sub-agent default to Opus	Pass model: "sonnet" explicitly every time
Forwarding the screenshot back to the main thread	Defeats the purpose. Sub-agent returns text only.