Fun AI Prompts That Actually Produce Something

Most fun ai prompts produce a chat window full of text that was amusing for the person who typed it and nobody else. The lists ranking for this search are long on that kind. Twenty prompts in five categories, fifty in ten, and in neither case a single sample output, which tells you the writers didn't keep anything either.

The fix is to sort by what you're left with. A fun prompt is worth keeping when it produces something that survives leaving the chat, a thing you can send to a friend, play with people in the same room, print, or look at. That's the filter for everything below. Fourteen prompts, shown in full, each with a note on what the role, input, constraints, and output shape are doing, because a fun prompt has the same anatomy as a working one and fails for the same reasons.

One of the ranking pages draws the line as functional prompts asking for information and fun prompts asking for a performance. Fair enough. The trouble with a performance is that it ends. I'd rather leave with a prop.

The Forwarding Test

The test I apply is whether the output still works when forwarded to someone who never saw the prompt. A roast of your to-do list fails it, since the recipient needs the list and the joke. A museum label for your friend's habit passes it, because the label is the whole object.

The column that matters most is the input one. Every prompt here needs something real from you, and that's not an accident. Fun prompts with no input produce the same output for everybody, which is why they stop being fun after the second person you show them to has already seen it.

Something To Send

These are for the group chat. The output is a finished object aimed at one person.

You write wall labels for a museum. Write the label for an exhibit called "[FRIEND'S NAME]'s [HABIT]", treating it as a significant artefact. Format: a title line, a date range, a materials line, then one paragraph of description in the flat, respectful voice of a museum. 80 to 110 words. No exclamation marks, no jokes that announce themselves. Use only these three true facts: [FACT 1]. [FACT 2]. [FACT 3].

The role is a museum writer, which buys a deadpan register without asking for one. The three facts are the input, and restricting it to those three is the constraint that stops the model inventing a fourth thing your friend never did. The format line gives it the shape of a real label so the output looks like something, not a paragraph.

Write the annual report for a group chat called "[NAME]". Members: [LIST]. Below are twelve things that happened this year, one per line. Produce a one-line mission statement, a "by the numbers" section with five statistics (clearly labelled as jokes), one award per member based only on the events listed, and an outlook for next year in exactly two sentences. Under 250 words. Events:
[PASTE EVENTS]

Corporate report as a frame does most of the work here. The "based only on the events listed" clause is the load-bearing one. Without it, the awards go generic and everyone gets "Most Likely To Bring Snacks".

Write a two-minute best man's speech in honour of my [OBJECT], which has been with me for [YEARS] years. Include one embarrassing anecdote built from this fact: [FACT]. Include one moment where the speaker gets emotional and recovers. End with a toast. Under 300 words. No puns on the object's name.

The "no puns on the name" constraint is there because that's the first thing the model reaches for and it's never good. Time limit expressed as speech length rather than word count gives it a pacing cue as well as a size.

Write a formal product recall notice for [FRIEND'S NAME], issued by the manufacturer. Affected units: all. The defect: [ONE TRUE FLAW, STATED KINDLY]. Include a model number, a "what to do if you own one of these" section, and a line about whether repair or replacement is offered. 120 to 160 words. Regulatory tone throughout.

Cruelty is the failure mode for this one, so the input line says "stated kindly" and the tone is regulatory, which keeps it dry. I'd send this to someone who'd send it back.

Something To Play

For a room. The model is the host, and the important design choice is giving it a rule that produces a dispute, because the dispute is the game.

We're playing a bluffing game with [N] people. Give me one real but obscure English word and its true dictionary definition, then write [N] plausible fake definitions in the same dictionary style, each 12 to 20 words. Number all of them, put the real one in a random position, and put the answer on its own final line starting with "ANSWER:" so I can cover it with my thumb. Don't use the word inside any definition. When I say "next", do another.

The output shape is what makes this playable. The answer on its own last line means the person holding the phone can hide it. Matching the length and style of the fakes to the real one is the constraint that stops the true definition standing out by being the only careful sentence.

Host a round of two truths and a lie about [TOPIC]. Give three one-sentence statements of matching length and specificity, so the lie doesn't stand out by being vaguer. After I guess, reveal the lie and give a one-line description of where each true statement could be checked. I keep score. Round 1.

One honest warning. The "true" statements will sometimes be wrong, because the model is confident about facts it doesn't have, and that turns out to be a second layer of the game rather than a problem. The check line is there so the room can argue about it.

We're writing a story together, five people in the room. Each person has given you one rule the story must obey. Rules: [RULE 1]. [RULE 2]. [RULE 3]. [RULE 4]. [RULE 5]. Write the story in under 200 words, obeying all five. Then list the rules and quote the exact sentence where each one is satisfied, so we can dispute it.

The dispute clause again. Five people arguing about whether "a character must lie about a vegetable" was really satisfied by a sentence about cabbage is the fun. The story itself is usually mediocre, and nobody minds.

Something About Yourself

The self-discovery genre is the biggest on this results page and most of it is horoscope. Ask what the model notices about you with nothing pasted and you'll get a warm paragraph about someone thoughtful who holds themselves to a high standard. That's everybody. The version that produces something is the version that starts from real material and is forbidden from adding to it.

Write a field guide entry for me as if I were a species, in the style of a bird guide. Use only the observations below and do not invent behaviour that isn't listed. Sections: common name, habitat, diet, call, distinguishing marks, behaviour, conservation status with a one-line justification. Under 180 words. Observations:
[PASTE 8 TO 10 TRUE THINGS ABOUT YOUR WEEK]

Bird guide is a good role because the register is precise and unimpressed. The "do not invent behaviour" constraint is the one people skip, and skipping it gives you the horoscope back with feathers.

Below is my calendar for last week, one event per line. Narrate it as a nature documentary in the voice of a calm, slightly concerned narrator. Reference only events in the list. Three paragraphs. End on Sunday.
[PASTE CALENDAR]

Same design. Real input, a rule against invention, a shape. The "slightly concerned" is my favourite two words in this whole piece, because it turns a Tuesday dentist appointment into a survival question.

Here are six mistakes I make repeatedly: [LIST]. Write a restaurant menu where each mistake is a dish. Each dish gets a name, a two-line description in menu language, and a price reflecting what it costs me in time or money (say which). Group into starters, mains, and desserts. No moralising, no advice.

The "no advice" line matters more than it looks. Left to itself the model will end a menu of your mistakes with a paragraph about growth. Nobody wants that on a menu.

A note on memory, since these prompts feed the model personal material. If your assistant has a memory feature switched on, it may pull in things from earlier chats you didn't paste. Google's help for the Gemini app says that with Memory on it may suggest topics or tools based on previous conversations, that you can ask "Did you use any info from past chats?" to check, and that it's switched off under Settings & help and then Personal Intelligence. Other assistants have their own version and their own switch. For a game with friends, the paste should be the whole input, so I'd turn memory off or use whatever temporary chat mode your tool offers before running these.

The deeper version of this, running a prompt with the input removed to see what's decoration, is in fun chatgpt prompts, and I won't repeat it here.

Something To Look At

Image prompts are where the "for selfies" searches go, and they pass the forwarding test almost automatically because an image needs no context. Two cautions before the prompts. Anyone whose face goes into the model should have said yes first, which is my rule rather than a citation. And any text you ask for in the image may come back misspelled, so keep names short and check them.

A movie poster for a trip that already happened. Title: "[TRIP NAME]". Tagline: "[ONE LINE]". The poster shows [ONE REAL PLACE FROM THE TRIP] at [TIME OF DAY], painted in the style of a 1970s adventure film poster, warm palette, hand-painted texture. Billing block at the bottom with these names: [NAMES]. Portrait, 3:4.

The one real place is the input. Without it the poster is a generic mountain. The 1970s hand-painted style is chosen because it forgives imperfect faces and text in a way a photographic style doesn't.

A collectible trading card for my [ANIMAL], [NAME]. The card shows [NAME] [DOING ONE REAL THING IT DOES], illustrated in a bold cel-shaded style with thick outlines and a foil-effect border. A stats box with four stats: Napping, Chaos, Appetite, Charm. The card's name at the top. Portrait, 3:4.

Four stat words is deliberate. More than that and the text rendering starts to fail. The real behaviour is what makes it your animal rather than a stock cat.

Restyle this photo of us as a Dutch Golden Age group portrait. Dark background, a single window light from the left, everyone in period dress, poses and expressions kept from the original photo, one dog added at the edge looking at the viewer. Keep faces recognisable. Landscape, 4:3.

Keeping poses and expressions from the original is the constraint that makes it funny, because the seventeenth-century clothes sit on top of a very modern slouch. Google's guide for its image model lists 4:3 as the preferred ratio for photography and 3:4 for more vertical content, which is why the group shot is landscape and the single-subject cards are portrait.

An archaeological site plan of my desk as if excavated in the year 2400. Top-down view, hand-drawn on grid paper, numbered finds with a legend: 1 mug, 2 tangle of cables, 3 [YOUR ITEM], 4 [YOUR ITEM]. Ink lines, light watercolour wash, museum drawing style. Square, 1:1.

The legend is where the text risk lives, so it's four short entries. The selfie-specific version of all this, meaning what gives a generated selfie away, is in ai selfie prompts.

What Falls Flat

The graveyard is consistent enough that I can describe it.

Prompts with no input, covered above. They produce the same thing for everyone.

"Be funny" with no constraint. The model's idea of funny without a frame is puns and exclamation marks. Every prompt above buys its humour from a register (museum, regulator, bird guide) and then forbids the obvious move.

Roasts. They read as a performance of bluntness and they're nearly always the same three jokes. The mechanics of why some humour prompts land and most don't are in funny ai prompts.

Prompts that open with a paragraph telling the model how brilliant it is. It doesn't help, and it uses the space where the input should have gone.

Long role-play setups. Four experts debating each other produce a panel that agrees with itself by paragraph three. The one-rule-each story works because the humans supply the disagreement.

Questions From The Results Page

What are some cool AI prompts? The ones whose effect doesn't wear off after one showing, which is a smaller set than the lists suggest. I've written about why the effects fade in cool ai prompts.

What's the most famous ChatGPT prompt? There's no official ranking, and anyone claiming a top ten is guessing. The answer Google pulls from the Reddit thread on this results page is the "tell me the most important 20% of a topic so I understand 80% of it" prompt, and the same commenter adds that it doesn't really deliver that. My own guess at the most-typed one is some version of "act as", which is a role line with nothing after it.

Are there fun AI prompts for friends? The first two sections, and the distinction is whether the friend is in the room or on the other end of a message. Different prompts for each.

What about fun prompts about yourself? Only with real input and a rule against invention. Otherwise it's a horoscope with your name on it.

What I'd Run First

The museum label, for someone you'll see this week. Three true facts, deadpan, under 110 words.

Send it without explanation. If they laugh, the prompt worked, and if they ask what it is, you've learned something about the forwarding test that no list of fifty prompts was going to tell you.