Instead of building custom apps, I prototyped two AI tools using existing platforms:
Claude Projects → Personal advisor (accesses my Notion journals)
ChatGPT Custom GPT → Personal nutritionist (static health data)
Initial assumption: “I need semantic search and vector databases”
Reality: File upload + smart instructions solved 80% of my needs. The remaining 20% became much more apparent with a live version that I could use.
Key insight: For applications that heavily rely on a chatbot interface, start with a CustomGPT / Claude Project.
After 2-4 weeks of daily use: Do the remaining limitations justify custom development?
In both these cases, I was able to start testing the idea quickly and observing where the limitations were. This helped determine a more robust feature set once I actually sat down to build the app. With the nutritionist for example, I decided to turn it into a web app so other friends and family could use it, I could add in weight tracking, and the ability to save insights that then were emailed back to me as a weekly report.
Prototype in existing AI platforms first. Let real usage drive technical decisions, not theoretical requirements. You’ll save months of development and build something you actually use.
The fastest way to build great AI tools is to start with tools that already exist.