DoomBreak

Scroll while it thinks. Stop when it’s done.

Comment

Inspiration

Every time I ask ChatGPT a “thinking-heavy” question, there’s a short pause. And every time, my muscle memory kicks in: I grab my phone “just to check something.” By the time the answer is ready, I’m still scrolling — and my attention is already gone.

I realised the problem wasn’t doomscrolling itself, but the in-between moments where distraction sneaks in automatically. So instead of fighting that habit, I decided to contain it.

That’s where DoomBreak came from.

What it does

DoomBreak detects when ChatGPT is generating a response and temporarily overlays a 3-panel, Shorts-style doomscroll wall. You get a quick, intentional brain break only while the AI is thinking.

The moment ChatGPT starts typing or finishes responding, the overlay automatically closes and brings you straight back to the conversation — no extra tabs, no “one more video.”

It also includes:

  • A daily auto-close streak to track how many times it saved you from infinite scroll
  • Sound controls (on purpose, not by accident)
  • Prompt-aware content filtering
  • A playful “discipline +1” close interaction

It’s not a site blocker. It’s a self-closing distraction.

How we built it

DoomBreak is a Chrome extension built primarily with:

  • JavaScript (content scripts + service worker)
  • DOM mutation observers to detect ChatGPT’s generation state
  • A custom video overlay system with panel-level playback control
  • Chrome local storage for preferences, streaks, and rewards

A big focus was making the detection robust on chatgpt.com, especially handling the difference between “thinking” and “typing” during streamed responses.

Challenges we ran into

  • Detecting typing reliably: ChatGPT’s DOM changes frequently, and streamed responses don’t always increase text length monotonically.
  • Autoplay & sound policies: Browsers really don’t want you playing audio unless the user explicitly interacts.
  • State edge cases: Manual closes, laggy generations, and rapid prompt submissions caused tricky race conditions.
  • Not overblocking: We didn’t want to block content permanently — just long enough to avoid accidental overuse.

Each of these forced us to rethink assumptions and make the logic more resilient.

Accomplishments that we're proud of

  • Getting thinking → typing → done detection to work reliably on chatgpt.com
  • Making the overlay fully self-closing without user friction
  • Turning a bad habit into a controlled, intentional interaction
  • Building something genuinely useful that I now use every day

Most importantly, it feels good to use — playful, not preachy.

What we learned

  • Small UX decisions matter more than big features
  • Fighting habits doesn’t work — designing around them does
  • Modern web apps change DOMs in unpredictable ways, and resilience beats elegance
  • A “personal” problem can still be a widely relatable one

What's next for DoomBreak

  • Support for other LLM platforms (Gemini, Claude, etc.)
  • Smarter prompt-to-content matching
  • Per-user analytics (private & local) on attention saved
  • Optional “focus mode” themes for calmer breaks
  • Packaging it so others can easily customise their own break style

DoomBreak started as a personal fix — next, it could become a healthier default for how we wait.

Built With

Share this project:

Updates