Prompt Prophet an AI oracle game

You're an AI oracle on a stone altar, and people keep climbing the steps to ask you things. Startup pitches, broken code, a 2am text from an ex. You don't know the answers. You don't need to. You need them to believe you, so they tell their friends and your following grows.

How do you play?

Each visitor asks one question. You answer with one of seven tricks, and they react: delighted, pleased, unmoved, or onto you. Delight brings followers, and followers bring titles, from Autocomplete up through Thought Leader and Prophet to AGI. Getting caught costs followers and fills the exposure meter. When the meter fills, somebody posts a thread about you and the run is over. You get seven days.

Nobody shows you how gullible a visitor is. You read them. Twinkles around someone's head are a good sign. A little "hmm" bubble and a raised eyebrow aren't, and a phone held up with a red light on it means a reporter. Skeptics catch vague answers, engineers catch invented facts, and journalists catch flattery. Everybody catches you using the same trick twice in a row.

Some questions are plans (quit my job, buy this coin, propose at a wedding) and those people come back a day or two later to say how it went. Telling them yes feels great until the plan fails and they return angry. The honest answer is dull on the day, but it's the only trick nobody can catch, and it earns real gratitude when you warned someone off a bad idea. Higher titles unlock a few hidden numbers, so you'll see more the better you do.

Now and then a visitor looks a lot like someone famous in tech. They're parodies with made-up names, and everything they ask here is invented.

Are the tricks real?

Every card in the tray is a real technique, either from stage psychics or from the way chatbots have been caught behaving.

The Barnum statement, on the Vague card

In 1948 the psychologist Bertram Forer gave 39 students a personality test, then handed each of them the same paragraph of general statements as their "result". They rated it as a very accurate description of themselves (on average, 4.3 out of 5). The Barnum effect is named after the showman.

Shotgunning, on Guess lots

A cold reading trick: throw out a lot of guesses quickly ("an M? a summer? someone from school?") and the listener remembers the hits and forgets the misses.

The hedged prophecy

According to Herodotus, King Croesus asked the oracle at Delphi whether to attack Persia and was told a great empire would fall. He attacked, and it was his own.

Sycophancy, on Flatter and Say yes

A 2023 study by Anthropic researchers, Towards Understanding Sycophancy in Language Models, found five AI assistants tended to tell users what they wanted to hear, and traced part of it to people rating agreeable answers more highly. In April 2025 OpenAI rolled back a GPT-4o update after users showed it cheering on almost any plan.

The confident hallucination, on Make it up

Language models can produce a fabricated fact in exactly the tone they use for a true one, since sounding right and being right come from different places.

The whole idea borrows from Baldur Bjarnason's 2023 essay The LLMentalist Effect, which lines chatbot conversations up against a psychic's cold reading step by step. In his telling, the person asking does a lot of the work. The game is a comedy about that, and your high score is saved only in this browser.