AI Prompts for Humor Writing That Are Actually Funny

I'm writing comedy about [TOPIC]. Don't write jokes. Instead, give me 10 specific,
weird, or annoying truths about this topic that most people have noticed but never
say out loud. The more specific and petty, the better. I'll find the jokes myself.

Picture a stand-up comedian at midnight, staring at a blank page before an open-mic set, knowing the premise is funny but unable to find the joke inside it. That's the moment where people reach for AI and get burned — because if you ask a model to "write a funny joke about airports," it hands you something that sounds like a joke but doesn't land. AI is genuinely bad at being funny on command. But the right ai prompts for humor writing don't ask it to be funny. They ask it to do the mechanical work around the funny — generating angles, finding incongruities, stress-testing a bit — so you can land the joke. This guide shows how.

Quick-Start (Copy This Right Now)

Don't ask for jokes. Ask for angles:

I'm writing comedy about [TOPIC]. Don't write jokes. Instead, give me 10 specific, weird, or annoying truths about this topic that most people have noticed but never say out loud. The more specific and petty, the better. I'll find the jokes myself.

What this does: It generates raw comedic material — the observations that jokes are built from — instead of finished jokes that fall flat. The "petty and specific" instruction is what separates usable material from generic setups.

⚡ Pro tip: Comedy lives in specificity. "Airplane food is bad" is not funny; "the way they hand you a napkin the exact size of a stamp" might be. Always push the model toward the hyper-specific detail, because the general version has been done to death and the specific one feels freshly observed.

Understanding the Variables

Specificity. The master dial, as with most writing, but doubly true for comedy. Vague is never funny. The joke is almost always in the precise, petty, over-observed detail.

Incongruity. Most jokes work by setting up one expectation and violating it. Understanding that lets you ask the model for the mechanism: "what's the expectation here, and what would surprisingly violate it?"

Truth. The best comedy is recognition — "it's funny because it's true." Ask for observations people privately agree with, and you're mining the richest vein.

Voice. Deadpan, absurdist, self-deprecating, sarcastic — these are different comedic engines. Naming yours keeps the model from defaulting to generic wackiness.

⚡ Pro tip: Ask the model to identify the "expectation" and the "violation" separately for any bit you're building. Once you can see a joke's machinery — setup expectation, surprising turn — you can build your own instead of hoping one appears. This is the closest thing to a joke formula that actually helps.

Step-by-Step: Building a Bit From a Premise

Step one — mine the premise (use the Quick-Start angle prompt above).

Step two — find the surprising connection:

From these observations about [TOPIC], which two are secretly about the same thing, or which one could be compared to something totally unrelated for a surprising angle? Give me the connection, not the joke.

What this does: Unexpected connections are where original comedy comes from. This finds the link; your phrasing makes it land.

Step three — write it yourself, then tighten:

Here's my joke draft. Don't rewrite it. Tell me: is the funniest word at the END of the sentence? If not, where should it go? And is there a word I can cut to make the punch faster?

What this does: Comedy is largely rhythm and word order — the funny word belongs last, and shorter is usually funnier. This audits your joke's mechanics without stealing the writing.

⚡ Pro tip: The funniest word goes at the end. "I bought a treadmill and now it holds my laundry" lands because "laundry" is the surprise, and it's last. Ask the model to check whether your punchword is buried in the middle of the sentence — moving it to the end fixes more flat jokes than anything else.

Pro-Level Variations

For a marketing copywriter adding wit to a brand: "Give me 10 observations about the annoying parts of [industry] that our product fixes. Keep them light and self-aware, never mean." Brand humor punches at shared frustrations, not at people.

For a speechwriter or best man: "Help me find the affectionate-roast angle on this person's habits — the loving version of teasing, not the cruel version." Warm comedy is a specific, learnable register.

For a newsletter writer who wants a lighter voice: "Where in this dry paragraph could a single unexpected aside make the reader smile without derailing the point?" Sometimes one well-placed line does more than a whole bit.

⚡ Pro tip: The "rule of three" is a genuine tool — list two normal things and one absurd one, and the absurd one lands harder for the contrast. "I packed sunscreen, a book, and my complete inability to relax." Ask the model to build you rule-of-three lists where item three is the joke.

Troubleshooting Common Issues

"The jokes are flat." You're asking for jokes. Ask for observations and connections instead, and write the jokes yourself.

"It's generic — I've heard these before." Not specific enough. Push relentlessly toward the petty, precise detail.

"It's trying too hard / too wacky." You didn't set a voice. Name your register — deadpan, dry, self-deprecating — and rein in the default zaniness.

⚠️ Common mistake: Trusting AI to judge whether something is funny. Models will confidently call a flat joke "hilarious" and rate your genuinely funny line as "confusing." They have no reliable sense of humor — they pattern-match to joke-shaped text. Use AI to generate raw material and audit mechanics like word order, but the final "is this funny?" call is always yours, ideally tested on a real human. A model's laugh means nothing.

A Prompt for Testing a Bit Before You Perform It

Comedy is the one form where you can't fully trust a page read — you need to know if it survives out loud. This prompt pressure-tests structure before you risk it on an audience:

Here's my bit. Don't tell me if it's funny — you can't judge that reliably. Instead, map its structure: where's the setup, where's the turn, and is there any wasted word between the setup and the punch that's slowing the timing? Just the mechanics.

What this does: It audits the timing architecture — setup, turn, punch — without asking the model to make the funniness call it's bad at. Tight timing is something you can verify mechanically even when funniness has to be tested on humans.

⚡ Pro tip: Dead words between the setup and the punchline kill timing. "So I was at the store yesterday, and there was this guy who was, like, standing there, and he..." buries the joke. Ask the model to find the shortest path from setup to punch, then perform that. Economy is most of comedic timing.

Your Turn

Take a topic you find funny and run the Quick-Start angle prompt. Pick the two pettiest, most specific observations. Then write three jokes yourself and run the word-order audit. You'll notice the AI never wrote a joke — it just cleared the underbrush so you could.

The angle-miner, the connection-finder, the word-order audit — these are the tools of a repeatable comedy process, and comedy is nothing if not repetition and iteration. Save your humor set in PromptABCD so that next midnight-before-an-open-mic, you start with ten specific angles instead of a blank page and a blinking cursor.