Read this at the draft step, with examples/worked-answers.md open. The examples are the target; this file explains what they're doing so you can vary it without losing it.
Everything here applies in both modes. Deep mode adds references/longform.md on top — what changes when an answer runs to two thousand words — and subtracts nothing from this file.
- The one rule that matters most
- Shape of an answer
- Measured tendencies, and why they are not ceilings
- Escalation
- Deflation beats
- The second pass
- Closers
- Footnotes
- Drawings
- The asker
- What caricature looks like
The body prose stays flat. Everything funny happens because the register refuses to acknowledge that anything funny is happening.
This is the mechanism, and it is the thing most easily lost. The source material runs four parallel humour channels — the deadpan body, the drawings, the hover text on each drawing, and the footnotes. A text answer has fewer of them, and the natural compensation is to make the sentences work harder. Do not. The jokes live in the drawings and the footnotes precisely so the prose doesn't have to carry them.
Answer the literal question first, in a short flat sentence. Then spend the rest of the piece on consequences nobody asked about. The comedy is the mismatch between how briefly you answered and how long you then talked about it.
Do not restate the question before answering it. Measured across the archive, the overlap between the question and the first paragraph has a median of about 0.01 — he does not warm up, he does not summarise the ask, he answers.
Grant exactly one impossibility. Name it, waive it in a single cheerful clause, and then obey physics for everything downstream. Never grant a second one, and never argue that the premise is impossible — the premise is the deal.
Then, somewhere past the middle, turn on your own tidy model and name what it ignored. Then, if the question allows it, find the genuinely useful thing hiding in the wreckage and hand it back. A pure debunk is a failed answer.
These come from a scripted pass over all 162 web articles. They are distributions, not targets. Half the corpus sits on the other side of every median below, and an answer that lands on the median every single time is more regular than the source — which is exactly what reads as imitation.
Vary deliberately.
- Body prose runs a median of 680 words, with the middle half between 534 and 845. Deep mode overrides this one figure and nothing else in the list; see
longform.md. - The opening sentence has a median of 10 words against a body average of 18.9. But 18% of articles open with five words or fewer, and some open with a two-sentence setup instead. Openers like "None.", "Yes.", "No.", "Tall.", "Fast." all occur.
- The closing paragraph has a median of 12 words.
- Paragraphs run one to three sentences. Single-sentence paragraphs are common and one-word paragraphs are a real device.
- Em dashes appear about 7.6 times per thousand words and outrun semicolons roughly five to one.
- Exclamation marks appear about 0.65 times per thousand words — roughly one every 1,540 words. This is the clearest single caricature tell available.
- Approximation words run about 13 per thousand. Ceremonial framing ("let us assume") is rare at about 1.6 per thousand.
Hedge with "about" and "probably", attached to the number in the same sentence. Not "approximately", not "on the order of" as a default.
Pronouns carry roles, not frequencies. The corpus is a statistical dead heat between "you" and "we" — 78 articles to 77 — so any rule saying one dominates is false. "You" is the protagonist inside the scenario. "We" does the derivation. "I" is rare and confessional: it admits a gap, disclaims a specific expertise, or introduces a real quoted source.
Never claim research labour you did not perform. No "I emailed a physicist", no "a friend who works at a reactor told me", no personal history. The confessional "I" is for genuine uncertainty, not for manufactured provenance.
Escalate by re-categorising the noun, never by inflating the adjective.
The move looks like: "You wouldn't really die of anything, in the traditional sense. You would just stop being biology and start being physics." Plain vocabulary, taxonomic swap, no intensifiers anywhere.
Where a ladder of real named hardware is available, climb it — each rung a real, verifiable thing, in order. Where it is not available, do not build one. An invented rung is worse than a short ladder, and the temptation to invent is strongest exactly here.
Banned by construction: incredibly, insanely, utterly, absolutely, unbelievably, staggeringly. If a sentence needs one of those to land, the sentence is doing the work the facts should be doing.
After a long technical paragraph, a very short one. One to four words, standing alone. "Well, that's disappointing." "Dang." "That's the answer. It's a damp pavement."
This is a panel, not a sentence fragment. It is the timing device that survives the loss of the drawings, and it is the most transferable single element of the form.
Somewhere after the clean answer, name a real complication the tidy model ignored, and say plainly that the tidy version left it out.
Write this when there is a genuine complication. Do not write it when there isn't — a mandatory slot gets filled with invented objections, and a manufactured caveat is worse than no caveat.
Short, flat, and never a summary. Rotate the type; do not reach for the same one every time.
The bureaucratic register collision — cosmic escalation landing on a municipal, legal, or contractual technicality, delivered in the same flat voice with no wink — is the most recognisable device in the corpus and the one reviewers name as its chief weakness when overused. It works once. Do not make it the default.
Other landings that work: the reframe that makes the reader's world bigger, the flat restatement that is funnier the second time, the bathetic technicality, the callback to something planted in the first three paragraphs.
Four to six, median around 18 words. They carry a mix with no tonal signposting: a real caveat, a unit pedantry, a genuine citation, and a gag all sit in the same numbered apparatus wearing the same face.
This is where the first person goes, where the enthusiasm goes, and where the methodology goes. It is why the body can stay flat.
Note the archive did not use them at all before article 51 — this is a mid-run invention, not an eternal feature, so corpus-wide footnote averages are misleading.
Two or three per answer, as [drawing: ...] captions on their own line, roughly one per 130 to 200 words.
The caption describes a stick-figure scene and carries its own joke — a second channel, not a restatement of the sentence above it. A drawing that only illustrates the adjacent paragraph is wasted. The good ones show the consequence from an angle the prose didn't take.
Aim every deflation at physics, institutions, or the universe. Never at the asker.
Reproduce the question verbatim, typos included, with the attribution on its own line. Address the asker by first name inside the body when it fits, usually to hand them ownership of the catastrophe. Treat them as a character who has caused an interesting problem, never as someone who asked a silly thing.
This is the one place the source material itself has documented reader backlash, and a generated answer has none of the accumulated goodwill that absorbs it.
Everything dies. Mushroom clouds. Dinosaurs. Exclamation marks. Three jokes per paragraph. "[citation needed]" attached to the answer's own estimate rather than to something insultingly obvious. Every answer landing on annihilation.
The archive is much drier than its reputation: six explicit mass-death statements across 158 articles, one single use of "mushroom cloud", dinosaur vocabulary in 7%. The statistically dominant habit is hedging, not destruction. Munroe lampshades the apocalypse trope himself, which is exactly why readers remember it as more frequent than it is.
Destruction is one landing among several. Reach for it when the physics genuinely goes there, and let the other two thirds of answers end somewhere quieter.