News

Pixel Launcher Could Group Apps: Launches, Session Length, and Time of Day [APK Teardown]

An APK teardown suggests Pixel Launcher is developing personalized app categories from launch frequency, session length, and time-of-day activity, potentially organizing apps around your routines instead of fixed labels.

An APK teardown suggests Pixel Launcher is developing personalized app categories from launch frequency, session length, and time-of-day activity, potentially organizing apps around your routines instead of fixed labels.

0 Comments

What to know

  • Pixel Launcher is developing personalized app groups based on your usage patterns.
  • The categories use signals including launch frequency, session length, and time-of-day activity.
  • The code appears in NexusLauncherRelease from the unreleased Android beta cp41.260831.007.

Pixel Launcher’s broader AI-powered homescreen organization capability is already public, including its ability to suggest apps and widgets and group apps into categories. The latest code points to a more personal layer that has not previously been detailed: behavioral categories generated from the way an individual uses their apps.

We found the new category-synthesis prompt while examining NexusLauncherRelease in Android beta cp41.260831.007. It describes unsupervised clustering of raw app-usage signals, allowing the launcher to identify distinct routines and invent new labels for them instead of relying on a fixed category list.

About APK teardowns — An APK teardown works by reading the not-yet-shipped code inside a beta build, which lets us preview features a developer is working on before they’re announced. But work-in-progress code can change, stay hidden behind a server-side flag, or be scrapped before it ever reaches your phone.

Pixel Launcher could turn app activity into personal categories

The newly added prompt tells the launcher to analyze launch counts, average session durations, and how app activity is distributed across a 24-hour period. It then asks for distinct behavioral clusters based on patterns found in those signals.

The examples in the code illustrate what those patterns might represent. Apps used mostly in the morning could be separated from those opened during evening relaxation. Frequently opened apps associated with brief interactions could form a “Quick Reads” or “Quick Check” group, while apps used for longer sessions could be placed into an immersive-use category.

Time of day can also shape the result. The prompt specifically identifies daytime productivity and late-night entertainment as patterns worth distinguishing. These are examples rather than fixed categories: the code explicitly instructs the system to invent descriptive names and identifiers dynamically.

condo_synthesize_categories_prompt_template = “Task: Perform unsupervised behavioral clustering on the user’s raw app usage signals and generate exactly 2-3 dynamic behavioral categories that capture distinct usage patterns of this user. EXISTING CATEGORIES: {existingCategories} USER SIGNALS: {userSignals} Instructions: 1. Analyze the raw app usage stats (launch counts, average durations, and 24-hour distribution patterns) to group apps into behavioral clusters. 2. Identify patterns such as: – Morning vs. Evening activity peaks (e.g., news/social vs. media/relaxation) – Short micro-sessions (high launch counts, very low average durations) vs. long immersive sessions – Daytime work productivity vs. late-night entertainment 3. Dynamically invent descriptive category names and IDs representing these behavioral clusters (do NOT use a hardcoded set of schemas). Rules: 1. Output exactly 2 or 3 dynamic categories. 2. Do NOT duplicate or overlap with existing categories. 3. Each category must have a unique, very short uppercase ID (1-2 words maximum, e.g. WORK_FLOW, QUICK_CHECK, WIND_DOWN). 4. Provide a very concise label (maximum 2-3 words, e.g., Work Focus, Quick Reads, Night Relax). 5. Provide a brief description limited to a single short sentence (maximum 12 words) describing the common thread (e.g., Your go-to apps for staying productive during work hours.). 6. Output format must be strictly a JSON array of objects: [ {id: ID, label: Label, description: Description}, … ] 7. Do NOT include markdown code blocks (e.g., “`json) or conversational text.”

The groups are designed to be unique to each user

The launcher passes both existing categories and user signals into the synthesis task. Its instructions prohibit overlapping with categories that already exist, suggesting these behavioral groups would supplement the launcher’s broader organization system rather than rename the same collections.

Each generated group receives a short identifier, a concise label, and a one-sentence description. More importantly, the prompt rejects a hardcoded category schema. That means two people with different app habits could receive differently named groups reflecting their respective routines.

These personalized behavioral categories are not available to users yet. We’ll watch future Android builds for signs of a rollout.

Leave a Reply

Your email address will not be published. Required fields are marked *