The Humor Project™

Class History

In past semesters, students worked in pods on shared Crackd applications. This semester is different. Instead of contributing to one internal codebase, you will individually build your own applications using Crackd's data model and the same kinds of AI tooling that we use in our development. The goal is to learn how to build an end-to-end, data-backed system that can (1) collect structured feedback from users, (2) support admin inspection and interaction, and (3) enable rapid prompt/pipeline experimentation for an LLM-driven content generator.

What is Crackd and What Are Humor Studies?

Crackd.ai is our web and mobile application that auto-generates captions for user-uploaded images using large language models. Crackd is also a research platform. We study what people actually find funny by collecting large quantities of votes on AI-generated captions. Our system is designed to leverage specific community context and "insider lore," especially Columbia student culture, because humor is contextual and audience-dependent.

That research component happens through humor studies. A humor study is a structured experiment that presents batches of captions to voters and records reactions (upvotes/downvotes) at scale. Each week, the class will run a study tied to a specific humor flavor, meaning a specific caption-generation strategy and prompt/pipeline configuration. On Fridays, a humor flavor will be triggered to generate captions. Those captions will be published to humorstudy.org, and students will vote on the results. Those votes do two things: they power our research analysis and they influence which captions feed into Crackd's "Top 100" feed for our real users.

How This Class Works

The fastest way to learn modern applied AI product engineering is to build real systems: frontends that read and write structured records, admin tools that support internal operations, and experimentation environments that turn subjective creative output into measurable signals. You will build three applications that mirror the three pillars of Crackd's internal workflow: public data collection, admin inspection, and pipeline experimentation.

Your work will use our existing Supabase database as the source of truth for shared entities (images, captions, votes, studies, and pipeline metadata). You will be given read-only access to the Crackd database so you can query it, learn the data model, and use it as the foundation for your products. For data you personally collect (such as votes captured through your version of a study site), you will write to a Supabase destination we provide for student-generated records. You are building on top of the same domain model that powers our platform.

Project Deliverables

In this class, you will be developing three separate applications. Each application is intentionally scoped to teach a different layer of the system and to reinforce how the pieces fit together.

The first application is a public-facing voting site for captions. This is your own version of humorstudy.org. It should load a study, display images and their candidate captions, and collect votes that represent "this caption is funny for this image." Your interface should make voting fast and intuitive. Your votes will be stored in the Supabase domain we provide so they become part of the shared research dataset.

The second application is an admin dashboard. An admin dashboard is essentially a frontend to the data in your database. For example, Crackd has a large internal Admin product because our AI systems require constant iteration, auditing, and debugging. You need to see what was generated, what prompts were used, how users reacted, and where your pipeline failed. Your admin dashboard will be a smaller version of this idea. It should allow you to browse and filter studies, captions, images, votes, etc. and support the types of questions you would ask while trying to improve a humor flavor. A good admin app is an interface that makes the dataset understandable. There is a lot of flexibility in design choice when it comes to your Admin area.

The third application is a prompt and pipeline testing environment for building your own humor model. This is your own version of matrix.almostcrackd.ai. Humor flavors are created through experimentation: you write prompt steps, run generations, inspect failures, adjust constraints, and repeat. This app should let you modify prompt steps and run test generations against known image sets to improve the flavor. Your goal is to make iteration easy.

Supabase

Supabase is our database platform. It provides a hosted Postgres database and additional services we use in development, including authentication, RLS, and an API layer. In this course, Supabase is the system of record. Every application you build will either read from it, write to it, or both. There is documentation on our website describing Crackd's domain model and the important tables you need to understand. You should treat the database schema as the contract between your applications and the Crackd ecosystem.

Development Environment

An IDE (Integrated Development Environment) is the primary tool developers use to write, navigate, and debug code. For this course, IntelliJ is the recommended IDE because it has strong tooling for backend and full-stack development, but it is not required. VS Code is also a strong option. What matters is that you pick an environment that allows YOU to move quickly and debug effectively. We do not recommend Vim for this course.

GitHub is a version control platform that developers use to store and manage code. A repo (repository) is a versioned project folder that contains your entire applications. You will commit changes as you build features so that you can track progress, revert mistakes, and collaborate easily when you need to share your code. You are expected to push code regularly and keep your repo in a working state.

Vercel is the hosting and development platform you will use to publish your applications. Deployment means taking your code and running it on an internet-accessible server so other people can use it through a URL. In practice, you will connect your repo to Vercel, and every push to a configured branch can trigger a new deployment. This makes it easy to test changes, share your work, and iterate without needing to manage servers manually.

AI CLI Tools

A CLI (Command Line Interface) is a text-based way to run commands and tools on your computer. In this course, you will use AI-powered CLI tools to accelerate development because that is the way our future is heading. These tools can generate boilerplate code, write components, and help debug errors. You may choose OpenAI CodeX, Google Gemini CLI, or Claude Code CLI.

The expectation is not that AI writes your entire project while you simply watch. The expectation is that you learn how to operate AI as a development partner. You should break work into small tasks, write precise instructions, validate outputs, and integrate changes. You are responsible for what you ship. AI is powerful, but it can also confidently make mistakes if you do not supervise it. You should still know how to read and debug code. A major learning outcome of this course is building the judgement to use AI tools effectively without losing correctness, clarity, or control.

What "Success" Looks Like In This Course

A strong project in this class is one where applications feel like real products, not just throwaway class projects. Your public voting app should make it easy to vote and should reliably record data. Your admin app should support investigation and analysis rather than dumping raw tables. Your prompt testing environment should make iteration fast and easy while keeping pipeline outputs traceable and comparable over time. Across all three, you should demonstrate that you understand the Crackd domain mode, that you can design interfaces that match the structure of the data, and that you can build systems that support experimentation. Additionally, a huge focus of the course should be completing weekly humor studies to provide us with high-quality data to contribute toward our humor research.