How we built Museum of Nonsense, a party game about real objects and dubious explanations
Inside Museum of Nonsense: real museum photographs, one shared multiplayer engine, private answers and optional AI polishing that leaves judging to the players.
A museum label gives you the answer before you have much chance to guess. Museum of Nonsense makes you wait.
An unusual real object appears on the shared screen. Everyone looks at it, then reaches for their phone. Depending on the mode, you guess what it was used for, choose from several explanations, or invent a convincing lie and hope your friends believe it.
The object is real. Its purpose comes from museum sources. Some of the explanations are complete nonsense.
That is the premise behind this game in the Playam collection. Building it meant solving two connected problems: finding objects worth arguing about, and making sure the software could keep up with the argument.
The collection is part of the implementation
An unfamiliar object is not automatically a good question. The photograph needs to be readable on a shared screen, the real purpose needs evidence, and the answer needs to make sense once it is revealed. An obscure catalogue entry with no useful explanation does not give us enough to work with.
The content workflow keeps museum records separate from the text written for the game. Source metadata and image rights sit alongside editorial answers, reveal explanations, translations and invented multiple-choice alternatives. That distinction matters: a plausible distractor must never become a historical fact by accident.
For new admissions, the review process retains the museum evidence and a digest of the imported image, so the review can be tied to the photograph we actually use. A local discovery catalogue remembers searches, candidates and rejection reasons. Finding something interesting starts a review; it does not publish a question.
The approved material becomes a prepared content pack, with photographs served locally by the game. There are no museum API requests during a round. A collection service being unavailable should not interrupt an evening halfway through a reveal.
Three modes, one set of rules
Museum of Nonsense has three round formats:
- Bluff: write an invented explanation, then vote among the submitted explanations and the real answer. You can earn points by finding the truth and by fooling other players.
- Written quiz: write what you think the object was used for. After the reveal, the group judges the answers.
- Multiple choice: choose from a prepared set of explanations.
Mixed Tour rotates through all three in balanced blocks. That gives a session variety without asking the host to start a different game each time.
Underneath, the formats share a deterministic TypeScript rules engine. It handles phases, deadlines, answers, votes, corrections and scores. The networking layer calls those rules rather than maintaining a second version of them.
This makes awkward cases easier to reason about. What happens when someone misses a vote? When two players write the same explanation? When a written answer deserves credit after discussion? Those are rules of the game, so they belong in the same place as scoring.
The engine also reconstructs state from the initial room and its accepted events. That gives us a way to replay what happened when investigating a disputed transition or score.
One room owns the game
The web app uses Next.js and React. Online rooms run through PartyServer on Cloudflare Durable Objects, with one Durable Object owning each four-letter room code.
A phone sends an action through the authenticated web API. The room server processes commands in sequence, validates them against the game rules, persists accepted changes and sends an updated view to each connected screen over a WebSocket.
Processing commands in sequence matters when several people submit at once. Each action is checked against the state left by the previous action, including when submissions arrive in the same millisecond.
Timers are also a server responsibility. Durable Object alarms handle deadlines, host-loss recovery and room expiry. An idle phone does not need to keep polling for changes or posting heartbeats to keep the game moving.
Room state and question reports expire 24 hours after the room was created. These rooms are temporary places to play together.
Keeping a secret means leaving it out
A bluffing game cannot send every browser the correct answer and rely on the interface to conceal it.
Each connected screen receives a view prepared for its identity and role. Before the reveal, that view omits the curated answer and revealing object titles. A player gets the information needed to take part; the shared display gets the public exhibition view.
Host control is separate too. The public joining QR code and private host-phone link serve different purposes. The display has its own identity and cannot issue game commands, even when opened in the host's browser.
That separation lets someone run the game from a phone while the television stays focused on the object, options, timer and standings.
Phones disconnect. The room has to cope.
Someone will lock their phone, switch apps or reload a tab. Reconnecting should bring them back to their seat.
The game uses a signed room cookie to keep browser identity stable throughout a session. Guests can play without creating an account, and signing in or out during a game does not replace that room identity. Reconnecting screens obtain fresh connection tickets and receive the current state.
Host loss needs its own rules. The room can pause and support takeover when the host disappears, so the other players have a recovery path.
Browser tests exercise full games across separate player contexts, including reconnecting, corrections and final scores. Rules tests cover deadlines, missing submissions, late joining and recovery. Those checks help us catch failures, while a real group trial remains a separate acceptance step.
AI can tidy a bluff. It does not award points.
An optional answer-polishing feature gives AI a deliberately small job: copy-edit a player's explanation and translate it into the room language.
The host enables it for the room, and players can still submit without it. The model receives the draft and target language. It does not receive the object, the true answer or player details. Its instructions are to preserve the player's claims and humour, without adding facts or making the answer more correct.
The original submission is locked first. Polishing has bounded retries and a short grace period; if the service fails or takes too long, the game can keep the original and continue.
Written answers are judged by the group after the reveal. Software tracks the decision and calculates the points. The discussion stays with the people around the screen.
One language across the room
The game supports Norwegian Bokmål, English and Swedish. The host selects the room language, and joining players adopt it across their phones and the shared display.
The content pack builder checks that required translations are present, that multiple-choice alternatives are distinct and that short answers fit their length limit. Museum credits and source records remain separate from translated game copy.
Those checks catch structural mistakes before play. Editorial review still has to decide whether an explanation is clear, fair and worth revealing.
What holds it together
The most useful implementation boundary is the one between the evidence and the game around it. Museum sources establish the real purpose. Players supply the bluffs. The server controls what each screen can see and when. The group settles the debatable answers.
That leaves room for a very serious photograph to produce a very unserious conversation.
Explore Museum of Nonsense on Playam, put an object on the shared screen and see who can make the least likely explanation sound authoritative.