When an AI product works with a reflective practice like tarot, the hardest design problem is not generating fluent text. It is defining what the system is allowed to claim.
I have been building TaroMind AI as a free guided tarot experience, and several product boundaries have proved more important than model cleverness.
For a Yes/No experience, the answer category should come from a transparent data structure. The model can explain a fixed answer, but it should not silently override it. This keeps the same card and orientation consistent across sessions and makes the behavior testable. The interface should frame cards as prompts for thought. Copy such as “What might this help you notice?” is safer and more useful than deterministic language. The product also needs visible limits around medical, legal, financial, and mental-health decisions.
A user's question can be deeply personal. It should not appear in URLs, analytics payloads, or public logs. In a browser flow, tab-scoped session state is a practical way to move the question between steps without leaking it into query strings.
Evergreen card guides, spread explanations, and simple local draws can remain available without authentication. The AI layer then adds interpretation for users who want it, while the core educational material stays crawlable and accessible.
Retrieval quality depends on the source material. If only part of the deck has detailed cultural context, the prompt must say so. The model should avoid inventing specific sources for cards whose data does not include them.
These choices make the system more predictable for users and easier to maintain. The working product is available at TaroMind AI, with public card references and free reflective tools.
The broader lesson applies beyond tarot: place important decisions in typed data and code, let the model explain within explicit limits, and keep sensitive user input out of observability systems.