Inspiration
In 2008, Nicholas Carr famously asked, "Is Google making us stupid?" Today, with Generative AI, that question is far more serious. As LLMs become incredibly smart, the cognitive friction to think has dropped to zero. We are witnessing a massive transition from cognitive offloading (using tools to assist us) to outright cognitive outsourcing (letting AI think for us).
This creates a hidden crisis: Cognitive Debt. By blindly copy-pasting answers without understanding the underlying logic, our skills decay. This issue hits close to home here, where AI is rapidly being integrated into school curriculums. We wanted to build a tool ensuring that as AI gets smarter, humans get smarter too. We still need to know what a good output looks like, just like we still need to understand multiplication even if we carry calculators in our pockets.
What it does
Codebt: Your Cognitive Debt protects your critical thinking skills while you use AI. Instead of letting AI do all the heavy lifting, it turns your AI workspace into a collaborative training ground.
- Real-Time Prompt Meter: As you type into ChatGPT, a dynamic meter scores your prompt locally based on depth, complexity, and structural rules. It encourages you to give smart constraints instead of lazy commands.
- Contextual Micro-Nudges: In-context warnings that stop users from mindless "vibe-coding" or copy-pasting.
- Timeline Effort Badges: Dynamic circular scores appended right next to user prompt bubbles in the chat log.
- Telemetry Side Panel: An isolated interface mapping your session scores, clear breakdowns of your operational habits, and showing a literal mathematical trace of the user's Cognitive Leverage Ratio.
How we built it
Codebt is a Chrome Extension engineered on top of the native Manifest V3 architecture using modular JavaScript, HTML5 canvas styles, and CSS variables.
GPT-5.6 acted as the architect. It took our raw idea to track prompt habits and mapped out the entire software structure. It helped us make the big design decisions, like building a Manifest V3 Chrome Extension, using standard HTML injection for the live meter, and building a native Side Panel to display user metrics. Once the structural blueprint was ready, it wrote a master plan to hand off to Codex.
Codex is the builder. It took the structural plans and instantly generated the clean code files by setting up the extension's background scripts, mathematical equations, and message routing. Codex did the heavy lifting by writing the source code step-by-step, debugging errors instantly, and adding new features smoothly without us getting lost managing multiple files.
Challenges we ran into
Targeting dynamic web structures can be exceptionally volatile. Building content scripts that inject custom UI directly into the ChatGPT canvas means our selectors could break whenever the target platform updates its frontend layout.
Another challenge, while utilizing AI models like Codex allowed us to do an incredible amount of heavy lifting rapidly, it introduced a counter-effect: moving at such speed meant we occasionally lost track of the broader codebase architecture, making it highly difficult to pinpoint exactly where unexpected errors were coming from.
Accomplishments that we're proud of
We are incredibly proud of bringing this project completely to life as a small team. We took a raw, abstract idea and built it into a fully working, highly functional tool that does exactly what we set out to do within a very tight timeline.
What we learned
We learned that introducing friction into a user interface can actually be a massive UX improvement if it serves a metacognitive purpose. In product design, we are usually obsessed with removing friction. Codebt taught us that "healthy friction" is essential for human education; slowing the user down for seconds to force them to think through their technical architecture preserves skill retention and alters how they interact with artificial intelligence.
What's next for Codebt
We want to expand out of the browser extension ecosystem, support wider developer environments and build an open-source Metacognitive SDK. This would allow educational institutions and school networks to integrate cognitive tracking directly into their internal portals. By doing so, schools can confidently roll out AI tools to students while receiving institutional-level telemetry data on whether AI is boosting active learning or quietly replacing student critical thinking.
Log in or sign up for Devpost to join the conversation.