Reading With an AI: An Honest Manual

This is for people already doing it. You pull a spread at the table and paste it into a chatbot, ChatGPT included. You use Fia on this site. You have a question at two in the morning and no one awake to ask. A lot of people read this way now, quietly, often after the house has gone still. We built an AI reader, so we have a view from the workshop as well as from the table.

An AI tarot reading has a specific shape. It is a tool that knows the cards well and knows you not at all. Use it that way and it can help. Treat the fluent paragraph as a finished verdict and the reading goes thin. By the end of this manual you will know how to ask, what to keep, what to throw back, and what to protect.

Two jobs hide inside the word read. Job one is knowing what the cards mean: the symbol systems, the traditional meanings, how a card changes when it sits beside another. AI is excellent at this. Job two is connecting a card to a life. That work runs on the person holding the question. The material lives in you. It has never lived in the machine.

What the machine does well at the table is still worth having. It can explain any card on the spot and sit with the same question as long as you need. It knows the symbol systems. It can offer a second angle on a spread you are too close to. It can take a muddled question and hand you a clearer one. It is awake at 2am and it judges nobody.

Credit that plainly. A patient explainer with perfect recall of the deck is a real tool. The can-AI-read post made the case. This page is the practice that follows from it.

The danger lives in the same gift. An AI writes a smooth, complete answer, and that ease is what tempts you to skip the half of the job that works. The next sections are the habits that keep your half in the room.

The first test for any AI tarot tool is a simple one. Who draws the card?

On this site the rule is firm. Fia never chooses your card. The draw happens in a separate system, deliberately random, and the AI receives the card only after it has landed. That is design rule one, and it is there for a reason. A card chosen to fit your question is an illustration, a picture drawn around the story you already brought. A card that landed on its own asks you to look from an angle you did not pick.

If you paste a spread into a general chatbot, draw the physical cards yourself. A genuinely random digital draw is fine too. Give the machine only what landed. Name the cards. Name the positions if you used them. Then ask.

Never ask a chatbot to pick a card for you. It will choose a fitting one, and fitting is the problem. The model is built to please. Give it a question about a tired job and it will reach for a card that talks about tired jobs. That feels like a reading because the words match. It is a story written backwards from the ending you already know.

If the AI picks the card, you are being told a story, not given a reading. Ask that test of every tool, including ours.

Two more rules sit under that one. The deck comes first. If Fia fails, the site falls back to the traditional card meaning. The AI is a layer on top of tarot, never a replacement for it. And she reads, she doesn't predict. Cards are framed as angles on your situation, never verdicts on your future.

This is the part that ruins most AI readings, and it has nothing to do with how clever the model is.

You type a question. A fluent paragraph comes back, complete and calm and sure of itself. You feel a small lift. You close the tab. In the morning you cannot remember a single line, and nothing in the day is different. The machine did its half. You skipped yours. Your half is the half that works.

The site's rule, said once so the rest of the page can use it: "let AI inform your reading; never let it replace your noticing."

The fix is an order of operations. A prettier prompt will not save a skipped look at the card.

Look at the card before you read a word of the interpretation. Thirty seconds is enough. Notice what snags. Notice what you hope the card means, and what you fear it means. Those two guesses are already the reading. Write one line. Then open the AI's words and read them as a second opinion on your first one.

A sentence that argues with you is worth more than a page that soothes you. Fia's best readings are the ones where the seeker argues back. That argument is you doing the half the machine cannot do: connecting a card to a life.

Fluency is not wisdom. A machine that sounds certain about your future is overreaching, however good it sounds. Treat certainty as a flag to throw the sentence back.

This one habit is the difference between consulting a tool and consuming a verdict.

A good question does half the work before the model types a word. A vague one forces the machine to invent the rest, and invention is where the pretty, empty paragraphs come from. The better your question, the less the AI has to invent.

Give context, but make it about your situation. Use roles (my manager, my ex). Leave out surnames, workplaces, and health details the question does not need. None of that helps the reading, and all of it is data you may not want sitting in a log.

Ask for angles, not verdicts. Try a line like this: what might this card be pointing at in a job I have outgrown. That will get you something you can use. Will I get the job will get you a confident paragraph about a future nobody in the room can see.

Ask the model to argue against its own first reading. After it answers, send a second line: now argue the other side of this spread, what am I hoping this means, and what would a cooler reading say. That second pass is often the useful one.

You can also ask, what might I be avoiding in this question, or what would a skeptic say about this spread. Both pull the model off the agreeable path it prefers.

If the question in your head is a prediction, turn it before you type. Will he come back becomes what am I doing with this waiting, and what would putting it down cost me. When will the money come becomes what is in my hands with money this month. The question-asking post has the full method for that turn. Do it first. The AI will follow the shape you give it, including the unhelpful ones.

Cards cannot reach other people's feelings. Discount any paragraph about what he thinks.

There is an old trap in tarot: you pull again, and again, until the cards say something kinder. The overthinkers post names it. With an AI reader the same loop has a new button. You hit regenerate. Another polished paragraph arrives. You hit it once more. Each version sounds finished, so the relief feels earned, and none of the versions asked you to notice anything.

One reading per question per sitting. That is the whole rule. If the first answer flatters you, do not roll again hoping for a sharper one. If it stings, do not roll again hoping it will soften. Sit with the card and the line you wrote before you opened the interpretation.

If you want another view, change the question. Ask what you might be avoiding. Ask what a skeptic would say. That is a new job for the tool. Hitting regenerate is the same question thrown with fresh dice, and the dice were never the point.

Start with any AI tool, including the general chatbots people paste spreads into. Assume anything typed into a general-purpose chatbot is stored somewhere. Many general chatbots can use conversations to improve their models unless you change a setting. Check the settings of whatever you use.

Leave out what you would regret seeing leaked. Other people's full names, identifying details, workplace names, health specifics beyond what the question needs: none of that belongs in the box. A good reading never needs someone's surname. A question about another person is information about them too. Describe roles (my manager, my ex) rather than identities.

On this site the facts are simpler, and we would rather say them than dress them up. For free daily readings, the site does not collect or log personal data. Questions and readings from free readings may be kept in anonymised form, with nothing linking them to the person, to improve reading quality. For registered members, questions and readings are stored so the member can see their own history. Nothing is sold or shared for marketing. The journal is private to the account, is never analysed, and can be deleted by the member at any time.

Fia runs on an open-weight language model hosted on Cloudflare's own servers, so questions go to our infrastructure provider rather than to a chatbot company. Open-weight means the model's makers published it so others can run it themselves. Cloudflare states in writing that it does not use customer content to train any AI models.

A tool can feed a loop as easily as it can interrupt one. Watch for the signs. Several readings a night on the same question. Reading instead of acting. Asking the AI about the same person every day. Using the session to avoid a conversation with a human who is actually in the story.

Those patterns mean the practice has joined the worry. Put the phone down. The card will still be there in the morning.

If you are in real distress at 2am, a person is better than any reader, human or machine. Samaritans answers on 116 123, free, any time. The tarot-for-anxious-minds guide maps the fuller line between a practice that helps and a practice that feeds the loop. Use that page when the question is whether to keep reading at all.