unstack

Inspiration:-

Many of us have experienced that sensation of having a whole lot to do and being unable to get started at all despite the approaching deadline. This inability is not due to slothfulness or lack of organization on the part of someone with ADHD and executive dysfunction. Instead, it is called task initiation friction, where intention fails to be carried out due to the nature of the task.

Tools like Notion, Todoist, and Google Tasks make the situation worse because they provide extensive lists that are cognitively overwhelming. Our goal was to develop something new, an application that would meet our users just where they are and give them one simple task to do within three minutes.

What it does:-

Unstack is an AI-driven task breakdown application that caters specifically to neurodiverse individuals dealing with executive dysfunction. All you do is inform it of the task you feel like you just cannot tackle at the moment, your energy levels, and your time availability.

The AI planner within Unstack will then break down your daunting task into the smallest possible set of practical, single action steps, all under 3 minutes in duration, all with verbs as first words, and all with almost zero activation energy to get started. No ambiguous commands, no decision-making, and no lengthy lists. Just: "Pick up 3 water bottles from your desk."

How we built it

-Frontend: Next.js 15 with the App Router, React 19, and Tailwind CSS v4 – designed for full keyboard support and screen reader support.
-AI Engine: API built on Google’s Gemini model using a custom-built multi-step prompt pipeline. The planner uses the LLM once to come up with an initial plan and then sends every single step to a deterministic validator which has strict constraints: no fuzzy verbs, no composite actions, max 3 minute action time. Steps that fail validation are automatically resubmitted to the model.
-Database: Supabase (PostgreSQL) database stores all sessions, steps, and interactions with feedback.
-Backend: Next.js serverless API routes are responsible for the whole planning process server-side.

Challenges we ran into

The most challenging thing about the project was making sure the AI gave specific instructions at all times. LLMs have a tendency to suggest general phrases such as "organize your notes" or "work on your essay." That is fine for neurotypical people. However, for people with ADHD, these instructions are utterly useless, precisely what triggers their paralysis.

This was overcome by creating a two-layered system where there is a very strict system prompt specifying 40+ approved concrete verbs, with the prohibition of vague instructions, along with a validation layer that checks the steps suggested by the language model against the rules and regenerates those not following them.

There were also difficulties associated with LLM providers' incompatibility and rate limits, making us create an independent provider abstraction through a special baseURL setting.

Accomplishments that we're proud of

-The validation/regeneration pipeline: The idea that AI-generated content undergoes verification and correction before it reaches the user is quite innovative.
-The design of the prompt: Our system prompt is the outcome of thoughtful consideration about ADHD and executive dysfunction problems. These rules are not random; they contain some actual cognitive science concepts (activation energy, single actions, concrete verbs).
-Accessibility: All interactive parts of the app have been keyboard- and screen-reader-friendly since Day 1. Every design decision in Unstack's UI was made with our users in mind. We deliberately avoided bright colors, complex navigation, heavy animations, and long lists, the very things that overwhelm neurodivergent users. A minimal, quiet interface isn't a lack of effort; it is the product. The design philosophy is: if the UI itself is calm, it gives the user's brain space to focus on the one task in front of them.
-The working end-to-end product: From an empty repo to an entire AI-driven database-based app within a hackathon period.

What we learned

We realized that AI prompting is as much a matter of limitation as one of capability. It's not that the planner from Unstack has an exceptionally intelligent LLM, rather it's that we've constrained it in such a way that it simply cannot get something wrong and not be found out.

We also developed a profound understanding of what it means to live with executive dysfunction on a daily basis. Designing this tool involved careful consideration of everything said, every interaction, every decision-making process within the interface, because each extra choice could mean a user leaving the app entirely.

What's next for unstack

-Session history & adaptive learning: Learn from past feedback ("too hard", "skipped") to personalize future step sizes and difficulty for each user.
-Sensory pacing themes: Visual countdowns, low-stimulation calm modes, and ambient soundscapes tailored to current energy levels.
-Campus integration: Partnerships with university ADHD support centers to offer Unstack as a structured study-aid tool.
-Mobile app: A native mobile experience optimized for on-the-go use — because task paralysis doesn't only happen at a desk.
-Break enforcement: Automatic Pomodoro-style break prompts baked into the step sequencer.