AI test prompts for Android apps: copy-paste tasks for your coding agent.
1. Smoke check on every build
Install the debug APK on my Phonebox phone and open it.
Sign in with the test account from TEST_ACCOUNTS.md, open each tab
in the bottom bar once, then sign out. Read the screen before every
action. Report any crash, error message or empty screen, quoting the
screen text. Park the phone when you finish.2. Exploratory session
Explore the checkout flow of the app on my Phonebox phone for 15 minutes.
Try what a hurried user would: go back in the middle of a step, leave
fields empty or very long, use a declined test card. Read the screen
before every action. List what felt broken or confusing, most serious
first, each with the steps to get there. Park the phone when you finish.3. Reproduce a bug report
Reproduce the bug in the ticket below on my Phonebox phone.
Install the latest debug APK, start a screen recording with
phonebox record, then follow the steps. Read the screen before every action.
If it reproduces, stop the recording, pull it to repro.mp4 and write the
exact steps with the screen text you saw. If it doesn't, say which step
went differently. Park the phone when you finish.
<paste the ticket here>The full workflow, with recording and handoff, is in reproduce Android bugs with AI.
4. Accessibility pass
Open each screen of the app on my Phonebox phone that I changed in this
branch. On every screen, read the elements with phonebox look and list
the clickable ones that have no text and no content description, with
their #id. Those are the buttons a screen reader can't name.It works because the agent reads each element's text and content description, not just pixels.
5. Before a release
Install the release candidate APK on my Phonebox phone over the current
version with --replace, so the existing test account stays signed in.
Run the journeys in docs/release-checklist.md one by one, reading the
screen before every action. Mark each pass or fail with what you saw.
Park the phone when you finish.What makes a prompt work
- Read before acting. Ask for a fresh read of the screen before every action, so the agent never taps from memory.
- Name the evidence. Ask for the screen text it saw, a recording, or both, so a report can be checked.
- Draw the line. Say what it must not do, such as real payments or messages to real people.
- Park at the end. A parked phone stops billing and keeps the test account signed in for the next run.
Questions
What is a good prompt for AI testing of an Android app?
Name the build to install, the screens or flows in scope, what to report and in what form, and what the agent must not do. Tell it to read the screen before every action. With Phonebox, the agent installs the APK on a cloud Android phone and runs the prompt there.
Do these prompts need a specific AI agent?
No. They work with any agent that can use an MCP server or run the Phonebox CLI: Claude Code, Codex, Cursor, Windsurf, Gemini CLI, Cline or your own.