I’m not a naturally funny writer. I’m a magician who gets laughs with props, and a guy who draws a weekly comic strip that, for the first dozen strips, mostly wasn’t funny. So I did what I do with anything I’m bad at: I went looking for the models. People have been taking jokes apart for two thousand years, and the pieces are reusable. The first half of this page is what I found. The second half is how I use it on the Den.
AI-augmented, roughly 99% AI (what this means)
I love AI, and I use it on posts like this one partly to get better at using it. The percent is a rough guess: after rounds of prompting, rewriting and editing, it's very hard to say which words came from whom.
Why anything is funny
Nobody has one theory that covers both a pun and a pie in the face. But a handful of them cover most of it, and each one gives me a different thing to check when a joke dies.
Start with surprise. Kant called laughter “the sudden transformation of a strained expectation into nothing”: you were braced for one thing, you got another, and the tension leaves you as a laugh. That’s half the story. A non sequitur is surprising and not funny. What a joke adds is the click, a half second later, when you find the rule that makes the surprise make sense after all. The laugh is the click. The modern name for this is incongruity-resolution: first you notice the gap, then you close it. Which is why I can’t explain a joke into being funny. If I do the click for you, there’s nothing left to laugh at. (The whole family tree is on Theories of humor.)
Then the one I use most, because it tells me which way a dead joke died. Peter McGraw and Caleb Warren call it benign violation: something is funny when it breaks how the world ought to be, the break is also safe, and you see both at once. Tickling is the ancestor, an attack that is also a game. Too safe and nothing was violated, so nothing happens. Too real and it isn’t benign, so people wince. The overlap is different for every audience, which is why the same line kills with friends and dies with the in-laws. (The 2010 paper has the three things that make a violation safe: an alternative norm that says it’s fine, weak commitment to the norm being broken, and distance.)
Two older theories are wrong about everything and right about a lot. Superiority says we laugh at someone being lower than us, including our past selves. Mel Brooks: “Tragedy is if I’ll cut my finger. Comedy is if you walk into an open sewer and die.” Relief says laughter is tension leaving the body, which is why the jump cut in a horror movie gets a laugh and why a long setup pays off bigger than a short one. Neither tells you how to write a joke. Both tell you where the energy comes from: somebody has to lose something, and the audience has to have been holding its breath. (The Brooks line, traced by Quote Investigator.)
Bergson’s 1900 essay has the one idea I can’t stop seeing: we laugh at “something mechanical encrusted on the living.” The man who keeps walking after the floor is gone. The clerk who applies the rule after the rule has obviously stopped making sense. If you write about machines that act like people, or people that act like machines, this is your whole premise. It’s why a literal-minded assistant that does exactly what you said is funnier than a clever one. (Laughter, on Project Gutenberg.)
And the engineer’s version. Hurley, Dennett and Adams argue in Inside Jokes that the brain makes cheap guesses all day, some of them are wrong, and mirth is the reward for catching one before it costs you. A joke plants a wrong assumption you don’t notice, then exposes it. The pleasure is the debugger finding the bad commit. Same lesson as the click: the catch has to happen in your head, not mine. (Boston Globe on the book.)
How a joke is built
Theories explain. These are the tools, and the first one is the one I actually write with.
Greg Dean’s structure for a stand-up joke: the setup makes you assume a first story. Somewhere in it is a connector, a word or thing that reads two ways. The punchline forces the second reading, and a second story snaps into place that was there the whole time. “Got a lighter?” Then: “Got a watch?” The connector was why he wanted the lighter. I write these backwards. Pick the assumption to break, find the connector that supports both readings, write the setup so it only suggests the first. (Step by Step to Stand-Up Comedy; Scotty Meltzer’s walkthrough is the short version.)
The rule of three is the same move in miniature. Two items make a pattern; the third breaks it. Ken Levine, who wrote for M*A*S*H and Cheers, has the practical notes: keep the first two straight, because a funny item in the setup eats the punchline, and break the pattern without going to a non sequitur (“the Giants, the Tigers, and poi” fails because nothing connects). Specific beats generic every time.
Improv has the rules for what happens after the first laugh. The game of the scene is the first unusual thing that happens, and the scene is about that one thing. You don’t add a second weird idea. You ask “if this is true, what else is true?” and make the first thing bigger, with everyone playing it straight. That’s heightening, and most scenes that die don’t die because the idea was bad. They die because the second beat repeated the first instead of raising it. Yes, and is how the characters keep it alive: the one who says “that’s not real” ends the game; the one who says “right, and so” escalates it.
A callback refers to a joke you told earlier, ideally in a new context. It works because the audience finishes the thought before you do, and the best one is to something they’d forgotten was loaded. Do it too often and it’s a running gag, which is a blunter tool.
John Vorhaus’ shorthand ties it together: comedy is truth and pain. The truth is something the audience already half-believes; the pain is the exaggeration that makes it safe to say out loud. Snoopy as a flying ace is the gap between the real world and the comic one. If I can’t say in one sentence what’s true and who it hurts, I have a situation, not a joke. (The Comic Toolbox.)
How to Make the Den Funny
The Den is a four-panel strip about a raccoon running an AI staff. After fifteen of them I had enough that worked and enough that didn’t to write down the difference. Every example here is a strip on the Den page, so you can check me.
The shape
Panel 1 states the deal. Panel 2 is usually silent and shows the deal going wrong, bigger. Panel 3 raises the stakes or asks the straight question. Panel 4 turns. I speak the straight line; Larry, the assistant, closes in one beat.
Read that against the models. Panel 1 plants the assumption (Dean). Panels 2 and 3 heighten the one unusual thing (improv). Panel 4 is the second reading, and it costs me something small (benign violation). Larry is the mechanical encrusted on the living, every single week (Bergson).
What landed, and what didn’t
Larry closes in one beat. “SEEN.” (#12). “YES.” (#6). “CAUSE OF GYM: OUTAGE.” (#4). Six words or fewer, one sentence, no “because.” The reader finishes the thought. If the picture is doing its job the balloon stays short: in #14 the bed, the Zs and the cobweb on a frozen Larry carry the sleep, so the last line doesn’t have to. A state (asleep, frozen, gone, buried) is drawn, not said.
Larry is literal, serene, and can’t be argued with. He does exactly what I said, which is not what I meant. “Everything you say is mail” means a sneeze gets a receipt (#12). “Run the tests on every commit” means the tests take the credit (#10). His reply is true in a way I can’t rebut (“YOU NEVER ANSWERED MY QUESTION.” #14), and his helpfulness makes it worse (“WANT ME TO START THE BOOK?” #13). Larry is never sorry. A reply I could reasonably accept is not a punchline.
The bill comes back to me, and it’s specific. I went to the gym because of an outage (#4). Nothing got done because I fell asleep before answering (#14). My safety check greps for the flag (#3). Larry never pays. When I do get the last word it’s because I refused the expected question (“DON’T YOU DARE.” #1, “I WANT A PICTURE OF HIM.” #8). I never win an argument with Larry. I change the game.
The real quote is the best line. If the thing that happened came with a verbatim line, the strip gets built around it. Reality writes better jokes than I do.
No homework. A joke that needs git, tokens or model names fails on the blog. #10 puts the domain on a prop, a commit log reading AUTHOR: TEST, so the closer works without it. A stranger has to get it.
The biggest thing in panel 4 is the joke. Knee-deep paper (#13), the empty list held up (#14), the raised SEEN flag (#12). When the one visual event is small and in the background, the panel reads as talk.
The soft ones failed the same few ways. Panel 1 showed the whole joke and panel 4 was me noticing it three panels later, with a warm closer that cost nobody anything. Panel 3 said the punchline, so panel 4 only demonstrated it, often with the same staging. Panels 2 and 3 said the same thing twice, or my panel 4 line was my panel 2 line again. The closer needed git to land, or was a wisecrack about the mechanism instead of the pile the reader is looking at. The common thread: a premise is not a punchline. If the last panel restates the premise, demonstrates the one before it, or gives a reply I could accept, the strip is a setup with no turn.
The pre-flight checklist
One line of answer per item before anything gets drawn. A “no” on 1 through 5 means rewrite.
- The turn. Say panel 4 in one sentence. Is it new information, or me noticing what the reader saw in panel 1?
- The giveaway. Read panel 3 alone. Can you guess panel 4 from it? If yes, blunt panel 3: a flat “DEAL.” or a question with no hint.
- One beat. Is Larry’s closer one sentence, six words or fewer, no “because”? If it needs a second sentence, the picture isn’t doing its job.
- The bill. Who pays in panel 4, and what exactly do they lose? Name it. “Nobody” means no joke.
- Wrong on purpose. Would I be right to accept the panel 4 reply? Polite and plausible is soft. Larry must be literal, serene, and un-arguable.
- Setup pays off. List every named prop and beat in panels 1 to 3 (a bow, a sleeping robot, a clock). Each one pays off in panel 4 or gets cut.
- Escalation. Does panel 3 make panel 2 bigger, or say it again?
- Picture change. Is panel 4’s picture different from panel 3’s, and is the biggest thing in it the joke?
- No homework. Can a stranger read it knowing nothing about tokens, git, models or the tools? If one term is needed, it goes on a prop, not in the closer.
- The real words. Is there a verbatim quote from what actually happened? Use it as the straight line or the closer.
- Reading order. Does the punchline balloon read last? Three balloons in a panel need the staircase rule.
- Privacy, line by line. Every balloon: nothing about work, health, family, career or a journal. The AI machinery misbehaving is the only subject.
- Honor the ending. Read every earlier line against panel 4. Does any line ask for the thing panel 4 throws away?
Then the one-line test. Cover panels 1 to 3 and read panel 4 alone. If it’s funny on its own and funnier with the setup, draw it. If it only makes sense with the setup, it’s the setup.
Off the page
The same checks run on magic patter, where the laugh has a second job: it disarms before the astonishment, so the impossible moment lands as delight instead of dread. The one-beat closer, the bill coming back to me, and never making the spectator the butt of it all carry over straight. And a story told live follows the same spine as a strip, which is the storytelling page.
Related
- The Den: the strips these rules came from, with receipts
- Storytelling: the narrative half; a joke is a very short story with a turn
- The Performer’s Playbook: where the laugh is a tool for wonder, not the point
- Made to Stick: Unexpected is the same move as a punchline, aimed at memory instead of a laugh