Logic puzzles for ChatGPT, Claude & Gemini
Updated
Deduction with one provable answer, sealed before release — the AI checks your reasoning instead of grading it kindly. Free, no signup. Each game is a ready-to-play prompt (we call it a game card) that you open in your own AI assistant.
An AI logic puzzle is a deduction problem — who owns what, who went on when — with exactly one answer that follows from the clues. The solution is sealed under a SHA-256 hash before release, so the assistant checks your reasoning against a fixed answer instead of agreeing with whatever you propose.
New to AI games? Read the best games to play with ChatGPT, Claude & Gemini.
| Game | What it is | Players | Length | Ages |
|---|---|---|---|---|
| The Ring of Five | A sealed logic grid: five bells, twelve true remarks, fifteen cells, and three smudges before the board is spoiled. | 1 | 20–30m | All ages |
| The Chairman's Book | A sealed ordering puzzle: seven variety turns, ten enquiries you pay for in minutes, and one submission the court never grades. | 1 | 30–45m | 12+ |
Can ChatGPT make a logic puzzle with one real answer?
A logic puzzle is worthless if the answer can move. Ask an assistant to invent one and it will produce clues that look rigorous and a solution that does not quite follow, because it wrote the clues forwards without ever solving them. Push back on its answer and it will often agree with you instead, which is the opposite of deduction.
These puzzles are built solution-first. The answer and a salt are written into a sealed file and hashed before the card ships, and every clue is generated from that answer. The assistant referees against the file: it can tell you a deduction is wrong, and it cannot be talked out of a fact.
Each puzzle is checked by exhaustive search before release, to confirm the answer is uniquely determined by the clues and reachable without guessing — no second arrangement satisfies them, and no clue is decorative. The Chairman's Book adds a second constraint on top: its evidence costs more minutes than the morning holds, so the puzzle is also a question of which testimony is worth buying. When you finish you read the reveal and recompute its hash at /verify/, which proves the answer you were checked against is the one committed before release.
Questions about logic puzzles
Is there exactly one solution?
Yes, and that is verified before the card ships by solving it from the clues alone and confirming every other arrangement is excluded. A puzzle whose clues admit two consistent grids, or which needs a guess to finish, does not ship.
Can the AI just give me the answer?
It holds the sealed grid so it can referee, and the card tells it to withhold the solution until you solve it or ask to be shown. Ask for a hint and it points at the next clue that constrains something, rather than filling in a cell. Whatever it eventually shows you, the hash check afterwards proves the answer was fixed in advance.
Do I need paper?
For The Ring of Five, no — the assistant prints your working grid every turn, so what you have inked and what you have eliminated stays on screen. The Chairman's Book is deliberately harder: it prints the enquiries you have bought, never a grid, and the court does not tell you how close you are, so the ordering is yours to hold or to write down. Either way, say save and your position travels with the code.
If you like logic puzzles
- Murder mysteries
Whodunits and interrogations whose solution is sealed before you start — the culprit can't change mid-game, and you check the seal after.
Blackfile №1: The Vanishing at Larchmont HallBlackfile №2: A Death at the Wyke HydroBlackfile №3: The Long Exposure at HarrowfellThe Third Carol: A Christmas Murder Mystery GameUnder Caution: An Interrogation GameNo Ghost Speaks: A Halloween Murder Mystery Game
- Riddles & lateral thinking puzzles
Riddles and yes/no mysteries whose answers are sealed before you guess — the AI rules from the file, not from whatever sounds plausible.
Twelve Chimes: Halloween RiddlesNothing Uncanny: Lateral Thinking Puzzles
- Chess puzzles & trainers
Fixed-answer puzzles and honest trainers — what AI is actually good at.