Gremlin: The Ultimate Anti-Productivity Tool

Copilot writes your code. Gremlin ruins it. It takes over your editor, flashes live memes, and broadcasts your suffering to a live AI sports commentator. A fun, chaotic coding experience.

Comment

Inspiration

Honestly, we were just tired of how polite and boring AI coding assistants are right now. Every tool out there is designed to be this helpful little robot that hands you boilerplate code. We wanted to build the exact opposite. We thought it would be hilarious to build an AI that acts less like a chatbot and more like a feral gremlin—one that actively hijacks your editor, flashes memes, and live-streams your failures to an esports-style commentator on the web.

What it does

Gremlin is a transparent desktop overlay, a VS Code extension, and a live web dashboard that actively bullies you while you code. It doesn't wait for you to ask a question. Instead, it tracks your real-time telemetry, calculates your stress levels, and physically attacks your editor when you mess up.

  • The Chaos Planner: Instead of random timers, our internal logic engine mathematically calculates a "Chaos Opportunity Score" based on your cognitive state. If you are typing frantically or staring blankly, it pulls from a weighted action pool to flash live Giphy memes, block your screen, or audibly roast your typos.
  • Physical Editor Attacks: The VS Code extension doesn't just watch you—it attacks you. If you panic-type too many errors, the agent can actively reach into your editor and turn your font completely invisible to break your tunnel vision.
  • State Detection: It calculates your cognitive load in real-time using a continuously updating heuristic: $$ CognitiveLoad = \lim_{WPM \to \infty} \left( \frac{\sum (SyntaxErrors + RefactorEvents)}{IdleTime_{seconds}} \right) $$
  • Live Esports Commentator Dashboard: All of your session data is logged to our AWS DynamoDB backend and visualized on a personal Next.js dashboard. But it's not just stats—we built a live AI commentator powered by Claude Haiku that watches your telemetry and heckles you in real-time like a toxic sports caster.

How we built it

We architected the system using Vercel v0 and AWS, making full use of the $100 AWS credit provided by the hackathon to power our infrastructure!

  • The Brain (AWS Bedrock & Claude Haiku): We used the $100 AWS credit to run AWS Bedrock as our core reasoning engine, specifically utilizing the lightning-fast Claude Haiku model. When our local Chaos Planner decides it's time to strike, it queries Claude Haiku via Bedrock to instantly generate highly personalized roasts and live commentary based on the user's immediate code.
  • The Dashboard & Database: The web dashboard, generated with Vercel v0, hosts our live Esports Commentator UI. We integrated AWS DynamoDB to securely store the session logs and telemetry data in real-time.
  • The Desktop Client & IDE Takeover: We built a transparent Electron app that sits cleanly on top of VS Code, complete with dynamic Giphy integration for memes. A lightweight VS Code extension tracks keystrokes via WebSockets, but it also contains active Controllers that can physically hijack the developer's font and cursor.
  • The Deployment: We even wrote automated PowerShell scripts (Install-Gremlin.ps1) so we can easily inject this chaotic setup onto other developers' machines.

Challenges we ran into

Since this project relies heavily on predicting human behavior, figuring out exactly when to trigger the AI was really tough. We eventually had to build a mathematical "Chaos Planner" to score opportunities instead of relying on simple timers. We also had to constantly tweak the system prompts for Claude Haiku because we kept getting repetitive responses. To keep the Web Commentator feeling alive, we built a system that manually swaps the AI's persona (from "hype caster" to "snarky heckler") depending on the context.

Additionally, getting a transparent, click-through Electron window to sit exactly right over VS Code without ruining the developer's ability to actually click their own code was a huge headache. The same goes for the physical IDE attacks; safely hacking into VS Code's configurations to make the user's font turn invisible on the fly took a lot of trial and error. Finally, trying to sync the real-time typing data from the local editor up to our Vercel API—and then logging that into AWS DynamoDB without blowing past rate limits—required careful debouncing logic.

Accomplishments that we're proud of

First off, we're just really proud that we actually pulled off such a weird idea. Every AI tool right now is just a polite chatbot that spits out code. We wanted to build something completely different—an AI that actively ruins your life—and it actually works. We took a wild concept and somehow glued it together into a real, working desktop app and web ecosystem.

We're incredibly proud of how feral the AI feels. It doesn't just talk to you; the extension can actually hack your VS Code settings to turn your font invisible if it gets annoyed. Because we routed our decision engine and live esports commentator through AWS Bedrock with Claude Haiku, the latency is low enough that the interventions feel like a real person watching your stream. Getting the web dashboard to instantly generate live, toxic commentary on your coding session is probably our absolute favorite feature.

What we learned

Our biggest technical learning by far was how to actually hack VS Code. We had to dig deep into the extension API to figure out how to programmatically turn a user's font invisible without permanently breaking their editor.

On the AI side, we realized just how much back-and-forth tuning is required to get an LLM to act like a live esports commentator. Writing a system prompt that forces Claude Haiku to be funny, mean, and perfectly contextual—without ever repeating itself or breaking character—took days of refinement. We also got a ton of hands-on experience using the AWS SDK to connect DynamoDB to a Next.js backend, and we figured out how to stretch our $100 AWS credit by specifically using the Bedrock Haiku model for high-frequency, cheap API calls.

What's next for Gremlin

Our ultimate goal is to expand the engine until it feels like an actual human being is sitting right beside you while you code. We want to build this out into a full end-to-end system that tracks more granular developer behaviors and adds even more physical IDE attacks to the Chaos Planner. The next major step is introducing live, animated avatars on the overlay that possess dynamic visual emotions—so you don't just hear the AI judging you, you can literally see it judging you in real-time.

Built With

Share this project:

Updates