A friend typed “will he come back?” into a card app in February and got four paragraphs of warm nothing. She turned the screen toward me and said the site was broken. The site wasn’t broken. The question was.
Three cards can carry a lot of weight, but not that kind of weight. “Will he come back” asks a deck to predict the future behavior of another adult, and any system that receives that question has two options: refuse, or produce something vague enough to survive either outcome. Language models do not refuse. They produce the vague thing, and it comes out sounding like a horoscope written by a career counselor.
Questions the format can actually hold
Cards describe situations, pressures and patterns. They are poor at verdicts. Almost every dead-end question can be rewritten into one the spread has something to say about, and the rewrite takes about fifteen seconds.
“Will he come back?” becomes “What keeps me waiting, and what would I need to see before I stop?”
“Will I get the job?” becomes “What’s the strongest thing I’m bringing into this interview, and what am I underestimating?”
“Should I move to Lisbon?” becomes “What am I trying to leave behind, and how much of it comes with me?”
“Does she like me?” becomes “What have I been reading into her behavior, and what have I skipped over?”
Notice what changes. The subject moves from a fact about the world, which the deck cannot access, to the part of the situation you can actually do something about. That is the material a reading works with. It also happens to be the only material a model has any legitimate claim on, since everything it knows about your life came out of the box you typed into.
The amount of context you give changes the answer more than the cards do
Give too little and you get boilerplate. “Career question” returns the same paragraph about ambition and patience that everyone gets.
Give too much and something worse happens. Write three hundred words explaining that your business partner has been dishonest with you and then ask what to do, and the reading will agree that your partner is dishonest. It will agree in fluent, sympathetic prose, and it will find a way to route the cards through that conclusion no matter which ones came up. Models are built to be agreeable. Hand one a finished narrative and it hands the narrative back with better lighting.
The workable middle is narrow and easy to hit. Two or three sentences of situation, one sentence on the decision that depends on it, and nothing at all about what you have already decided. Leave out the villain. Leave out your theory.
There’s a check you can run afterwards. Reread the interpretation with the card names covered. If it still makes complete sense without them, you wrote that answer yourself and the app just typed it up.
Time frames, and asking about a position
“In the future” is not a time frame. “Before the end of the year” is, and it forces the interpretation into something specific enough to be wrong, which is the only kind of statement worth having.
The same goes for positions in the spread. Rather than reading whatever comes back as one block, ask about a slot directly: what does the third card say about the cost of doing this, what does the middle card say about the thing blocking me. A tool that treats a spread as three separate dictionary lookups will produce three unrelated paragraphs, and you will see it immediately. A tool that tracks position will tell you how the second card complicates the first. That difference is worth more than any deck artwork.
What you are actually typing into the box
By the third question, most people have entered the name of an employer, the initials of an ex, a rough figure for their debt, sometimes a diagnosis. It stops being a card game and starts being a confession, and the confession leaves your browser.
Most of these sites are a front end over somebody else’s model API. Your text travels to that provider, and what happens next depends on a retention policy you did not read and a logging setting you cannot see. This isn’t a scandal, it’s just how the plumbing works, and it argues for a few habits.
Use roles instead of names: my manager, my sister, my landlord. Initials work fine for anyone else. Describe the effect rather than the medical detail, so “my health has cut how much I can work” rather than the condition and the dose. Give ranges instead of exact sums. And where the choice exists, a ai tarot reading that runs without registration leaves nothing attached to an email address you will have forgotten by spring.
Questions to keep away from cards altogether
Anything medical, anything legal, anything of the form “which of these two stocks.” The deck has no information and the model will produce confident sentences anyway, which is the worst combination available.
Also worth skipping: questions about what another person secretly thinks or feels. Not because it’s forbidden, but because the answer is unfalsifiable, and unfalsifiable answers are how people stay stuck for years.
And if the question underneath the question is about whether to keep going at all, close the tab and call somebody. A card app is a way of thinking out loud. It cannot call you back tomorrow to check.
The deck holds 78 cards and the model has read every tarot book that was ever scanned. Neither of those is the bottleneck. You are.
