Guides
How to come up with an AI project, not just a wrapper over ChatGPT
Why wrappers lose, how a real product differs from a prompt with an interface, and how to find an idea that ten other teams won't bring.
The most common AI project at a hackathon looks like this: an input field, a button, a request to a chat model, and the response on the screen. Technically it works, but the jury sees dozens of such wrappers and almost always gives them a low score. Let's break down how a real AI project differs from a wrapper and how to come up with an idea that stands out.
Why wrappers lose
A wrapper doesn't solve the problem; it passes it on to the user. Its value is exactly the same as that of the model itself, which anyone already has access to. Over a weekend, dozens of such projects are submitted, they all look alike, and they are impossible to remember. The jury doesn't evaluate the fact that you called an AI, but what you built around it.
What turns a wrapper into a product
- Its own task. Not a 'chat with AI', but a specific pain point of a specific person, where the model's response is only a part of the solution.
- Its own data. The project works with something the bare model doesn't have: your documents, database, context, integration.
- Work before and after the model. You collect, verify, and connect something with other systems, rather than just forwarding a request.
- Reliability. A product figures out what to do when the model makes a mistake, whereas a wrapper just shows the error to the user.
A simple test: if you remove the beautiful interface, is there still value that isn't in ChatGPT itself? If not, it's a wrapper.
How to find an idea that stands out
- Start from the pain point, not the technology. First, a problem you know from the inside, then a solution where AI is appropriate.
- Narrow it down to a single scenario and a single audience. A narrow, useful tool beats a universal 'AI for everything'.
- Add something that cannot be Googled in a minute: your own data, your own logic, your own context.
- Check the event tracks. On the hackathon page in Stavleak, tracks and prizes are detailed, and they often explicitly state which tasks are valued, rather than 'bring us another chat'.
An idea is only half the battle; next, you need to build and showcase it. How to bring a project to a working demo with AI and not drown in generated code is covered in the guide on vibecoding, and you can find a suitable event in the catalog.