Say what you mean: why plain English beats brittle selectors

Every flaky mobile test has the same root cause. Somewhere, a script names an element by how the app was built (a resource id, an XPath, an accessibility label three layers deep) and one day the app is built a little differently. The element is still right there on the screen. The test can’t find it.

AppClaw takes the opposite starting point. You describe the element the way a person would point at it, and the agent resolves that intent against whatever is actually on the screen right now.

The problem with a path

A selector like //android.widget.Button[@resource-id="com.app:id/btn_primary"] is a path through a tree that only exists because someone wired it that way. It carries no meaning about what the button does. Rename the id, wrap it in one more container, ship an A/B variant, and the path breaks, even though nothing a user would notice has changed.

Worse, paths are ambiguous in the ways that matter and precise in the ways that don’t. A screen with a “Login” button and a “Log in” text link has two perfectly valid matches for the word login, and the selector that disambiguates them is exactly the one most likely to break.

Describe it the way you’d point at it

AppClaw lets you name the target by its relationship to something stable:

tap login button below the password field

There’s no id in that sentence, and there doesn’t need to be. The agent reads the screen, finds the password field as an anchor, and resolves the button geometrically, the one below it, the same way you would if you were looking over someone’s shoulder. below, above, left of, right of, near, and within all work as spatial anchors.

When the accessibility tree is missing entirely (Canvas, Flutter, games, custom-rendered UI) the same instruction falls through to AI vision on the raw screenshot. The instruction doesn’t change. Only how the agent sees does.

Intent survives redesigns

The quiet payoff is durability. A plain-English step encodes the goal, and goals change far less often than layouts. “Turn on Wi-Fi” is true across every Android skin and iOS version; the selector that accomplishes it is different on each. When you author against intent, a redesign that keeps the feature keeps your test.

This is the whole thesis behind AppClaw: automation is shifting from brittle scripts to agents that understand. You bring the intent. The agent handles the path.