This is a submission for the Weekend Challenge: Dog Days Edition.
My submission is PawDOS 95, a fake desktop operating system built for dogs. Built to boot like a 1995 PC, it tracks squirrels, monitors the mailman, manages treats, and tries to explain humans with a translator.
The main feature is DogPrompt. It looks like a DOS terminal, but it accepts normal sentences. Gemini does more than answer in character. It can request approved actions that affect the desktop itself.
PawDOS has a working window manager, start menu, taskbar, dialogs, notifications, synthesized sounds, random desktop events, and 9 applications.
DogPrompt combines about 25 instant local commands with Gemini-powered responses. Other apps include Squirrel Tracker 2.0, HUMAN.EXE, Treat Manager, Mailman Monitor, My Kennel, My Documents, Notepad, and Settings.
The app also works on mobile.
Try PawDOS at pawdos.dog.
Open DogPrompt and type something like:
emergency i see a squirrel outside
Gemini can respond, open Squirrel Tracker, add a radar target, show a warning, and trigger the bark driver. The terminal finishes typing first, then PawDOS carries out the approved actions.
The source is available at github.com/pinkpixel-dev/pawdos.
PawDOS uses React 19, TypeScript, Vite, and Zustand. A single Cloudflare Worker serves the app and handles the Gemini requests.
DogPrompt uses Gemini function calling. The Worker gives Gemini eight possible actions, including opening an app, spawning a squirrel, barking, changing the dog's mood, and entering nap mode. Every returned action passes through an allowlist validator before the browser can run it. Invalid names, values, and arguments are dropped.
HUMAN.EXE, squirrel dossiers, and hardware diagnostics use structured responses. The Worker validates those fields before React renders them.
I am submitting PawDOS for Use of Google AI.
Gemini is connected to the operating system instead of sitting in a separate chat box. Function calls become validated desktop actions, structured responses power three utility panels, and conversation chaining gives DogPrompt short-term context.