Separating the Draw from the Text
The program fixes the cards, orientations, and positions first. The language model then generates text using your confirmed question, context, and card definitions. The model is not responsible for choosing cards to reach a preferred conclusion, nor should it alter the cards already drawn.
If the interpretation service is unavailable, the site displays its status. You can still look up public card meanings directly; those texts are not personalized answers generated for your specific question.
Your Actual Words Outweigh the Category
Categories simply help you find an entry point. If you start from a reconciliation topic but end up writing "I don't want to get back together, I just want to move on," the reading must address moving on. If your title mentions accepting a job offer but your context states you have no written offer yet, the text must not fabricate a new job for you.
Your question, provided facts, spread positions, and card symbols must work together. Common romantic or career clichés should never override your actual words.
What Makes a Card Meaning Useful
Consider the question: "I keep practicing, so why am I not improving?" With the Eight of Pentacles, the reading can examine whether practice includes concrete feedback loops. With the Ten of Wands, focus shifts to whether sheer overload prevents the practice from being absorbed. Both suggest adjustments, but for different reasons under different observable conditions.
If every card simply produces "believe in yourself and be brave," the unique role of the card has been lost.
What Structural Checks Catch
The site verifies whether an interpretation contains required elements and checks that card names, orientations, and positions match the actual draw. If the response is incomplete or swapped cards, it registers a failure and preserves the original draw.
These checks cannot guarantee that every sentence fits your life. Checking whether the content truly answers your question remains something only you can do against your real situation.
How to Tell if the Reading Truly Responds to You
Check first whether the opening paragraph addresses your primary question directly. Next, verify whether the quoted context came from you. Finally, see if the suggested next step connects to this specific card in this specific position. If your clear constraints are ignored, or if unmentioned meetings, promises, and inner motives suddenly appear, do not accept the reading just because the prose flows smoothly.
Unsuitable parts should simply be recognized as inaccurate or inapplicable. You do not need to reshape your reality to fit the text.
Data Boundaries and Service Limits
Your question, background context, and drawn cards are sent to our server first. Email addresses and Taiwan-format phone numbers in your question or context are masked automatically before processing, storage, and transmission to OpenAI or TypeSafe; other phone formats are not, so please do not enter personal data. The cards are fixed before text generation begins; the AI never redraws cards.
When sending requests to OpenAI, we configure the service not to retain conversations. However, OpenAI still handles content under its own data policy; this does not mean the entire process leaves no record. TypeSafe also receives your question and context to determine whether tarot is suitable; if you follow up from a previous reading, a brief conclusion from that reading may be included.
We retain request logs long-term for operational analysis and abuse prevention. These logs include your question, context, card spread, timestamp, full IP address, country, network provider, browser type, anonymous visitor identifier, and AI token usage. Full IP addresses are stored only in our server logs and are never shared with OpenAI or TypeSafe.
Please do not enter full names, contact information, or identifying details about yourself or others. For retention periods, data handling, and deletion inquiries, see Data & Privacy.