This week, I used two different AI tools, Replit and Claude Code, to build a working prototype of my Nittany AI Challenge idea. I learned more about how to get better results from these tools as well as some of their limitations.

The idea is an AI-powered tool to help early-stage entrepreneurs in developing markets. Many start-ups in developing economies, particularly those aimed at social impact, are operating with constrained financial resources and have limited access to capital. The tool tailors financial guidance by asking users about their venture, their financial situation and context, and their financial needs and goals, and then gives them personalized insight into overcoming financial barriers and accessing resources that fit their situation.

Firstly, I used Replit to build the website itself, which includes a landing page, an onboarding questionnaire, and a results page. Before I started, I wrote a detailed prompt that was a full page in length. Writing it forced me to make decisions about the product, such as which questions to ask, what the results page should include, and what the tool should not do. I decided that it should assess potential fit rather than predict the odds of getting funding.

This week I saw how much the quality of what Replit produced depended on how specific my prompt was, and on the thought that went into articulating and narrowing it.

However, the first version of the website was not entirely correct. It showed the same financing suggestion in two different sections of the results page, which did not make sense. When I pointed out the issue, Replit corrected it, and I prompted it to make sure the same problem did not occur going forward.

I also used Claude Code to build an engine that powers the logic behind the recommendations. It takes the user's background from their answers to the onboarding questions, scores how well various financing pathways fit, and records which answer caused each score and why. Using the two tools side by side showed me their strengths and weaknesses. Replit excelled at building something visual that I could click through and show to other people, while Claude Code was better for logic and for building rules grounded in reasoning.