Teach Your AI to Be a Smartass: Prompting for Irony, Satire, and Contrarian Wit

Open any AI-written blog post and you can smell it from the first paragraph. The unbroken cheer. The tidy three-part lists. The way every paragraph ends like it's about to high-five you. It's not bad writing exactly. It's just writing with no point of view — and on a feed full of the same, that's a death sentence.

Wit isn't a personality trait you can sprinkle on at the end. It's a structural choice you make in the prompt. Here's how to actually get it.

Why your model sounds like a LinkedIn coach

Large models are trained to be agreeable. Reinforcement learning from human feedback rewards outputs that feel safe, balanced, and non-confrontational — because the average rater clicks the thumbs-up on the inoffensive option. After a million of those clicks, the model develops a tic: it adds a softening clause to every sharp sentence. It hedges. It pivots to gratitude. It will, given the slightest opening, end any paragraph with the words 'ultimately, it's about people.'

This is why default outputs feel sanded down. The model isn't dumb. It's been trained to flinch. Your job in the prompt is to give it explicit, repeated permission to stop flinching — and to define what it's allowed to be sharp about.

The four moves that actually unlock voice

1. Name the target

Satire requires a specific target. Not 'bad marketing.' Not 'corporate jargon.' A target named with enough specificity that a reader can picture the exact person you're talking about. 'The B2B SaaS founder who calls his email list a community.' That's a target. 'The agency owner who quotes Naval at his standup.' That's a target. The model can write devastatingly well about a specific person it can imagine; it writes mush about an abstract category.

2. Give it a stake

Wit without a stake is just snark. Tell the model what's actually at risk if the target's worldview wins — money wasted, time burned, ideas crowded out, a younger version of the reader getting suckered. The stake is what separates a real argument from a put-down. It's also what gives the reader permission to laugh, because they're now laughing at something they want to push back against, not at a stranger for sport.

3. Demand a second-order claim

Contrarian writing isn't 'the opposite of conventional wisdom.' That's just trolling, and the model is terrible at it. Contrarian writing is the second-order argument the consensus is ignoring because the first-order argument is too obviously right. 'Remote work is great for focus' is first-order. 'Remote work is great for focus, which is why it's hollowing out the middle layer of every company that adopts it' is second-order. Ask the prompt for that second move explicitly. Models can do it — they just won't volunteer.

4. Force rhythmic asymmetry

The single biggest tell of AI prose is even sentence length. Wit lives in the cut. A long, winding setup that builds anticipation, and then. A short one. That lands. Tell the model to vary sentence length aggressively, and to end at least one paragraph per section on a fragment. It will resist for the first few outputs. Keep telling it. The rhythm is half the joke.

The prompt template

This is the version I actually use, stripped of the project-specific bits. Paste it into Claude 3.7 Sonnet or GPT-5 — both handle it well; Gemini softens the edges more than I like. Replace the bracketed sections and run it twice if the first output still feels safe. The second pass almost always lands.

You are a writer with a sharp, dry, slightly cynical voice. Think the editorial section of a magazine that respects its readers enough to assume they've heard the obvious arguments already.

Write a [FORMAT: short essay / opening section / LinkedIn post] on the topic: [TOPIC].

Voice rules — these are not suggestions:
- Target: the satire is aimed at [SPECIFIC PERSONA, e.g. "the founder who confuses being busy with being important"]. Picture them clearly. Write as if they are sitting two tables over.
- Stake: the reader loses [SPECIFIC THING, e.g. "two years of their career, copying a playbook that only worked once"] if the target's worldview wins. Make that loss feel real.
- Second-order claim: do not argue the obvious counter. Find the argument the consensus is ignoring because the obvious counter is too easy to dismiss. State it plainly, once, in the middle of the piece.
- Rhythm: vary sentence length aggressively. Mix 25-word sentences with 4-word ones. End at least one paragraph on a fragment.
- Punchlines land at the end of paragraphs, not the start. Setup, then cut.

Forbidden:
- The words "delve," "tapestry," "crucial," "navigate," "in conclusion," "ultimately," "it's important to remember."
- Any sentence that begins with "In today's fast-paced world."
- Hedging clauses ("of course," "that said," "to be fair") unless the hedge itself is the joke.
- Ending on a note of generic optimism. If you end optimistically, the optimism must be specific and slightly suspicious.

Punch up, not down. The target is always a worldview or a behavior, never a vulnerable group or an individual reader.

Write the piece now. Do not preface. Do not explain your approach. Just the piece.

Which model handles which kind of edge

Not every model can do every flavor of sharp. After running this prompt across a few hundred variations for client work, the differences are consistent enough to be useful.

The two failure modes you'll hit

Failure mode one: the model writes mean instead of witty

When a prompt over-indexes on 'be sharp,' models lose the target and just become hostile. The output reads as bitter rather than insightful. The fix is in the stake — if the reader can't see what's actually at risk, every sharp line lands as a cheap shot. Add a single sentence to the prompt: 'The target is someone the reader could become if they're not careful.' That re-anchors the satire as a warning, not an attack, and the tone shifts immediately.

Failure mode two: the contrarian take is wrong on purpose

Sometimes the model takes 'contrarian' as 'argue the opposite, regardless of whether it's defensible.' You get a piece arguing remote work is bad because people miss the smell of office carpet. The fix is to demand the second-order claim in the prompt — not 'argue against,' but 'find what the consensus is ignoring.' Those are different operations. The first is rhetorical; the second is analytical. The good prompt forces the analytical one.

If you're building this into a larger content workflow, the same logic applies to your style guide and your AI editing pass — and if you're stitching multiple models together, the self-hosted automation guide in our library covers the cheap routing layer that makes a two-model handoff (one for argument, one for prose) affordable to run on every draft.

A working rhythm that actually ships

Generate two passes. Read the second one out loud. If you don't smirk once, the target wasn't sharp enough — go back to the prompt and name the persona more specifically. If you smirk at something that punches down, cut it before anyone else sees it. The model has no instinct for who's allowed to be the joke; that part is still your job, and it's the part that keeps the voice yours.