A mystery works in AI roleplay when the solution is decided before play starts, the clues are fair, and every suspect is hiding something. Keep the cast of suspects small, state each clue plainly when it appears, and question suspects one at a time.

Decide the solution before play starts

An AI writes each reply from what is in the conversation and the setup at that moment. If nobody has decided who the culprit is, the answer can shift halfway through the story. Whoever writes the mystery should fix who did it, how, and why in the adventure setup or the character definitions. If you play a mystery someone else created, that work has been done for you.

Play fair with clues

Golden Age detective writers treated a mystery as a contest with the reader. In 1929 Ronald Knox published his “Ten Commandments” of detective fiction, and one of them says the detective must declare any clue he discovers. The same rule works for AI roleplay: every clue needed for the solution should appear in the story before the reveal.

Use clues that are missing

In Arthur Conan Doyle’s “Silver Blaze” (1892), Holmes directs the police “to the curious incident of the dog in the night-time”. The dog did nothing, which means the intruder was someone it knew.

Something that should have happened and didn’t makes a good clue for AI roleplay. A character can mention it in passing (“Odd. The gate was locked when I arrived.”) without giving the answer away, and it’s up to the player to notice.

Trap a small group of suspects

Agatha Christie returned to the closed circle again and again: guests stranded on an island in And Then There Were None (1939), passengers on a snowbound train in Murder on the Orient Express (1934). A trapped group keeps the list of suspects short and the pressure high.

Four to six suspects is a manageable number for an AI story. Questioning them together in a multi-character scene lets you catch them contradicting each other.

Give every suspect a secret

If only the culprit is hiding something, the mystery is solved the moment someone acts nervous. Give the innocent suspects secrets too, such as an affair, a debt, or a lie about where they were. They produce red herrings that feel natural, because those suspects really are lying, just about something else.

Let witnesses disagree

Wilkie Collins tells The Moonstone (1868) through several narrators, each of whom knows part of the truth and gets some of it wrong. Question suspects separately and compare their accounts. When two stories don’t match, ask each suspect about the other’s version.

Reason out loud

Edgar Allan Poe’s detective C. Auguste Dupin, in “The Murders in the Rue Morgue” (1841), solves the case by reasoning from small observed details. Write your own deductions into the story (“The ticket is dated tomorrow, so he planned to leave”) and the AI characters will react to them by confirming, objecting, or panicking.

Sherlock Holmes sums up the method in The Sign of the Four (1890): “How often have I said to you that when you have eliminated the impossible whatever remains, however improbable, must be the truth?”

Keep track of the evidence

Mysteries depend on details from early chapters, and those are the first things an AI can lose in a long story. When you find a clue, state it plainly once. Use the same name for each suspect and object throughout. At the start of a session, recap the case in a line or two: “The ticket, the missing ledger, and the captain’s lie about the tide.” See why AI characters forget.

Pay off what you plant

Anton Chekhov repeatedly advised young writers that a gun shown in the first act should go off later. A mystery also needs the reverse: whatever turns out to matter at the end should have appeared earlier.

Lunelle AI’s long-term memory keeps clues from early chapters available at the reveal, across sessions. See long-term memory.

How to run a mystery story with AI characters