Recent findings reveal that Google’s new Magic Pointer feature is written almost entirely as AI prompts and sends portions of your screen to Gemini for processing.

Google’s new laptops, expected to debut later this year, will feature an interface powered not just by code, but also by AI prompts sent to Google Gemini. The code also reveals something broader: how Google’s own engineers are now building product features entirely out of AI prompts rather than code.

A recent teardown by Android Authority shows how Google’s Magic Pointer functions. Unlike a traditional cursor, which triggers pre-written code when clicked, Magic Pointer works by sending text prompts to Google Gemini along with images grabbed from your screen. Google’s APK code is typically heavily obfuscated, making it intentionally difficult to read and understand, however, on this occasion, we can see the full plain text prompts sent to Gemini.

Google has not specified whether user data will be processed by an on-device Gemini Nano model or in the cloud. This is a key privacy question that remains unanswered.

Making Magic Pointer a standalone app, rather than part of the operating system, means users get to opt in as they wish. It won’t be forced on every Googlebook owner.

What The Magic Pointer Teardown Reveals

The teardown reveals some important findings about how Magic Pointer will work:

  • You must have the Google app configured as your default assistant.
  • Magic Pointer will be optional: you don’t need to use it to use a Googlebook
  • You can bring up Magic Pointer by wiggling your cursor or typing a keyboard shortcut
  • You can configure the sensitivity of the wiggle or turn it off altogether

These findings are backed up by the actual text prompts Gemini uses to power Magic Pointer.

Google’s Instructions To Gemini, Word For Word

Here’s the system prompt that defines the overall role and goal of the Magic Pointer. This steers the way Gemini will behave in all Magic Pointer interactions:

You are the intelligence behind the Magic Pointer for Android laptop OS feature.

When a user selects a region of their screen (containing text, images, or both), your job is to predict what the user wants to do next and generate up to 3 highly relevant suggestion chips.

These chips act as direct prompts into Gemini.

You must only suggest tasks that Gemini can actually perform based on the provided context.

Further prompts define four categories of action: understand, transform, ideate and execute, with Magic Pointer choosing the most relevant suggestion for what's on screen. This reveals something unexpected about how Google is now building software.

We’re seeing, for the first time, a user-interface tool constructed almost entirely through “prompt engineering” (the skillful construction and refinement of AI prompts) rather than traditional code. Google has set several guardrails, for example requesting gender-neutral language and “culturally appropriate” responses. However, these rules are enforced by simple AI prompts rather than strict code. This approach always leaves room for the AI to respond with something incorrect or inappropriate.

Upgrading Magic Pointer could be as simple as changing a prompt. It also allows for a great deal of user-customization should Google decide to allow it. But, this flexibility cuts both ways: If the company were to allow users to change these prompts, it could open up serious vulnerabilities such as prompt injection or users gaining unlimited free access to Gemini models to perform other tasks.

This offers a glimpse at how tomorrow’s laptop interfaces may be engineered through conversation rather than code. While we’re waiting, you can try out two official Magic Pointer demos for yourself, on Google AI Studio: Point and speak to Edit an image or Interact with Google Maps .

You can read the prompts in more detail in Android Authority ’s original report.