🪦 Digital Graveyard

Why I Vibecoded My Own Time Tracker

3 min read

Why I “Vibecoded” My Own Time Tracker

I’ve always struggled with losing time without knowing where it went. I tried existing desktop activity watchers, but they didn’t fit my life:

  • Too complex: Over-engineered for my simple needs.

  • Linux Hurdles: Many struggle to run consistently on GNOME + Wayland.

  • Unreliable: They would stop tracking without warning.

  • Android Gaps: Digital Wellbeing is great, but it’s a “walled garden”—no easy sync, export, or integration.

So, I decided to build my own solution using Antigravity (with Sonnet/Opus and Gemini 1.5 Pro as a backup). I’m not an Android dev, but I understood the logic well enough to “vibe” the code into existence.

The Breakthrough: Android + Tailscale

Despite having zero background in Android development and finding the built-in Android Studio AI agents frustrating, I pivoted. I started chatting directly with Gemini, moving code into Android Studio bit by bit.

The result? A functional backend-heavy app that tracks like Digital Wellbeing but stays under my control.

Then, I discovered Tailscale, which changed everything.

  • Personal VPC: It feels like having a private Virtual Private Cloud for just my devices.

  • No VPS Needed: I can access a website deployed on my laptop directly from my phone.

  • Generous Free Tier: 100 devices for free is plenty for my home/office setup.

My Architecture & Sync Flow

I currently manage three main nodes: a Personal Laptop (Ubuntu), an Office Laptop (Windows), and an Android Phone.

  1. The Master Node: My personal Ubuntu laptop acts as the “Source of Truth.”

  2. The Sync Strategy:

    • Android: Toggle Tailscale and sync/upsert data directly to the master.
    • Office: Sync data via Drive, then merge it into the master locally.

  3. AI Auto-Categorization: Since the raw data is messy, I built a workflow to auto-categorize activities using AI. I use Cline (or Antigravity) to run these workflows.

    • Models: I prefer Claude Opus or Gemini Pro for this logic.

    • Pro Tip: I avoid GPT-4o-mini/OSS models for code-heavy tasks; they tend to break the structure. I stick to the heavy hitters for accuracy.

Check out my prompt: If you’re using Claude Code, Cline, Roo Code, or Kilo Code, you can see how I handle the logic in .agents/workflows/auto-categorize.md.

Key Takeaways

  • Understand the “How,” not just the “Detail”: You don’t need to be a senior dev to build tools if you understand the architectural logic.

  • Tooling matters: Tailscale + modern LLMs (Sonnet/Gemini) turn a “wasted afternoon” into a productive deployment.

  • Vibecoding is real: It’s about maintaining the flow and using the right model for the right task.

Graph View