This is a small food-tracking and planning application in which a language model acts as a lightweight application layer over a focused interface and its data.
Open the food-planning experiment.
What I was exploring
The project tested how quickly useful personal software can be shaped when the interaction model is conversational but the product still owns the interface. The useful lesson was not “put a chatbot everywhere”; it was that a narrow context and clear UI can make model behavior easier to direct and evaluate.
Next engineering step
A production-quality version should enforce allergies, budget, and other safety-sensitive constraints with deterministic validation. Any saved change should also require an explicit user action. Those boundaries are part of the planned Meal Planning Copilot agent demo, which will be published only after it is testable.
Evidence
