Most habit-tracking apps are fundamentally broken. They act as digital cheerleaders, relying on the user's temporary motivation and the myth that habits are formed in 21 days. We wanted to build a system grounded in actual behavioral science.
HABIT@66 was inspired by a synthesis of three foundational texts:
Atomic Habits by James Clear (Systems over goals).
Tiny Habits by BJ Fogg (Motivation is a flaw; habits must take <30 seconds to start).
The Power of Habit by Charles Duhigg (The neurological loop of Cue -> Routine -> Reward).
Combined with Phillippa Lally’s definitive 2009 study proving that automaticity takes an average of 66 days, we set out to build an AI that doesn't motivate you, but surgically engineers your behavior.
What it does HABIT@66 is an AI-driven behavioral architecture platform. Instead of letting users input ambitious, doomed-to-fail goals (like "Run a marathon"), the platform's AI Architect ruthlessly deconstructs their goal into a micro-action taking under 30 seconds (e.g., "Put on running shoes").
The AI then interrogates the user to extract an "Anchor"—an existing, non-negotiable daily action (like pouring coffee or flushing a toilet) to act as the neurological cue. Once the system is locked, the user is dropped into a 66-day tracking dashboard. The platform strictly enforces the 30-second baseline for the first 22 days before automatically scaling the difficulty in phases.
How we built it Built by a two-person, cross-university team (Polytechnique Montréal and Concordia University) in just 8 hours, we utilized a modern AI-assisted workflow:
Frontend Generation: We used Lovable.ai to rapidly scaffold our React and Tailwind CSS components, focusing on a minimalist, dark-mode aesthetic to match the rigid, architectural brand of the app.
Backend & DB: We ripped out the mock state in Cursor and wired the application to a Supabase PostgreSQL database to handle user state, streaks, and the 66-day grid logic.
The AI Engine: We intentionally abandoned a slow, hallucination-prone RAG pipeline. Instead, we utilized In-Context Learning via the Claude Opus 4.6 API. We engineered a massive "Mega-Prompt" that acts as a strict state-machine, forcing the LLM to abandon conversational pleasantries, guide the user through a 3-step behavioral gauntlet, and return strict JSON payloads to update the frontend UI.
Challenges we ran into Scope Creep: Midway through the sprint, we realized building a RAG pipeline to search the books would consume our entire 8 hours and yield sluggish results. We had to pivot hard to the Mega-Prompt architecture to save the timeline.
Wrangling the LLM: Forcing an advanced model like Claude Opus to stop being "helpful and chatty" and instead act as a rigid, uncompromising architect was difficult. Ensuring it consistently outputted perfectly structured JSON without Markdown formatting to trigger our UI state changes required heavy prompt optimization.
The Integration Handoff: Bridging the gap between the static React code generated by UI tools and a live Supabase backend while fighting the clock tested our full-stack capabilities.
Accomplishments that we're proud of Shipping a complete, beautifully styled, full-stack application from absolute zero in under 8 hours.
Successfully programming an LLM to act as a definitive "state-machine," dynamically generating tailored Phase 2 and Phase 3 habit escalations based purely on a user's initial prompt.
Creating an AI application that actually solves a human problem with strict scientific frameworks, rather than just acting as a generic text wrapper.
What we learned Ruthless Prioritization: In a hackathon, an instantly responsive, hyper-focused UI is vastly superior to a bloated backend. Dropping the Calendar API and RAG pipeline was the best decision we made.
AI Meta-Coding: We learned the optimal flow for building software in 2026: use visual AI tools (Lovable) for high-speed UI scaffolding, and agentic IDEs (Cursor) for deep-tissue backend wiring and database integration.
Prompt Engineering is Software Engineering: Writing the logic for our Claude API call wasn't just writing text; it was writing strict compiling instructions for a neural network.
What's next for HABIT@66 Automated Escalations: Fully integrating the database triggers to dynamically shift the user's required habit from Phase 1 (<30 secs) to Phase 2 and Phase 3 as they cross the 22-day and 44-day thresholds.
Contextual Triggers: Implementing an SMS/WhatsApp integration to ping the user contextually based on their specific Anchor time, rather than a generic daily alarm.
Accountability Contracts: Adding a feature where users stake money on their 66-day streak, creating an artificial consequence for breaking the neuro-loop.
Log in or sign up for Devpost to join the conversation.