Un-Smoothing the Copy: Forcing AI to Write Sharp Openings and Conclusions That Don't Suck

Every piece of AI copy I've ever cleaned up had the same two tells. The intro warmed up for three sentences before saying anything. The ending restated the intro in a slightly different jacket. In between, the body was usually fine. The problem isn't the model. The problem is that nobody told it to stop being polite.

I've shipped a few hundred AI-assisted pieces in the last year — landing pages, newsletters, product one-pagers, the lot. The prompt scaffolding below is what I actually use. It's ugly. It works.

Why AI copy reads smooth in the bad way

Default model behavior is to hedge. Hedging reads as smooth because it's frictionless — there's nothing sharp to disagree with, so nothing snags. That's also why it's forgettable. A sharp opening makes a claim a reader could argue with. A sharp conclusion gives them something to do. Smooth copy does neither.

The fix isn't a better adjective list. It's structural. You have to ban the shapes the model reaches for when it's stalling.

The shapes you need to ban

• Rhetorical questions in the opening ("Ever wondered why...?")
• Tricolons — three-item lists used for rhythm rather than meaning ("faster, smarter, better")
• The "not just X, but Y" construction
• Any sentence starting with "In a world where..." or "In today's..."
• Conclusions that begin with "Ultimately," "In conclusion," or "At the end of the day"
• Em-dash sandwiches used to delay the verb

Banning shapes works better than banning words because the model can swap synonyms but can't easily swap structures. Tell it not to write tricolons and it has to actually pick one idea.

The opening: fix the first sentence, the rest follows

Almost every AI intro fails because the first sentence is a setup, not a claim. The model writes "Writing good copy is hard." Then it writes "That's especially true with AI tools." Then, finally, on sentence three, something useful. Cut the first two.

My rule: the first sentence has to be a sentence the reader could disagree with. "AI copy reads smooth because the model is hedging" is arguable. "AI is changing how we write" is not. The model defaults to the second. You have to demand the first.

Three openings that work

• The specific anecdote: a single concrete scene, no setup. ("Last Tuesday I deleted 400 words from a draft before adding a single one.")
• The contrarian claim: state the thing your reader assumes, then reject it in the same sentence.
• The number drop: a real, specific number that implies a story. Avoid round numbers — they read invented.

The conclusion: write it first or skip it

This is the trick that changed my output more than anything else. **Make the model write the conclusion before the body.** When the ending exists, the body stops padding to reach it. When it doesn't, the model writes a body that wanders, then improvises a conclusion that summarizes the wandering.

Better still: ban the conclusion entirely for short pieces. Anything under 800 words doesn't need one. End on the last real point. The reader knows the article is over because the article is over.

The prompt I actually use

Drop this in as a system prompt or prepend it to your draft request. It's long on purpose — every constraint is there because I watched a model violate it.

You are drafting copy for a publication that prizes specificity and rhythm over polish.

TOPIC: {{TOPIC}}
AUDIENCE: {{AUDIENCE}}
LENGTH: {{WORD_COUNT}} words, hard cap.
FORMAT: {{FORMAT - e.g. essay, landing page, newsletter}}

WRITE IN THIS ORDER:
1. First, draft the final sentence of the piece. It must be a concrete instruction, observation, or image — not a summary.
2. Then draft the opening sentence. It must be a claim the reader could disagree with. No setup, no warm-up, no question.
3. Then the body, working toward the ending you already wrote.

BANNED SHAPES (do not use any of these, ever):
- Rhetorical questions in the first paragraph
- Tricolons (three-item rhythmic lists like "faster, smarter, better")
- The construction "not just X, but Y" or any variant
- Opening with "In a world where," "In today's," "Imagine," or "Picture this"
- Closing with "Ultimately," "In conclusion," "At the end of the day," or "Remember,"
- Em-dash sandwiches used to delay the main verb of a sentence
- The words: delve, tapestry, crucial, leverage, robust, seamless, navigate (as a verb for non-physical things), unlock, harness, journey (as a metaphor)

RHYTHM REQUIREMENTS:
- Vary sentence length aggressively. At least three sentences under six words. At least one over thirty.
- No two consecutive sentences may start with the same word.
- No paragraph longer than four sentences.

SPECIFICITY REQUIREMENTS:
- Every claim about a number, time, or quantity must be specific. "A few minutes" is banned. "Seven minutes" is fine.
- At least one concrete proper noun (a tool, a place, a person, a brand) per 200 words.
- If you cannot meet the specificity requirement honestly, leave a [BRACKETED PLACEHOLDER] for me to fill in. Do not invent.

WHEN YOU FINISH:
Return the piece, then below it, list any banned shapes you were tempted to use and what you wrote instead. This is for my QA.

How to tweak it

For landing-page copy, add a line forcing the first paragraph to name the reader's exact situation in under 15 words. For newsletters, swap the "final sentence first" instruction for "write the subject line and the P.S. first" — those are the only two parts most readers see. For long-form essays over 1500 words, relax the paragraph-length cap to six sentences or the piece starts to feel staccato.

The QA list at the bottom is the secret weapon. It makes the model audit itself, and you'll learn which constraints your model cheats on. (Claude reaches for tricolons. GPT-4-class models love em-dash sandwiches. Gemini will use "crucial" until you ban it three times.)

Which model handles un-smoothing best

They all need the scaffolding. But they fail in different ways, and the prompt above performs differently depending on which one you point it at. Here's what I've seen across maybe 200 runs of similar prompts.

What this fixes and what it doesn't

Un-smoothing makes copy sound like a person wrote it. It does not make the copy accurate, well-argued, or worth reading. If the underlying thinking is thin, you'll just get a punchier version of thin thinking. The prompt is a finishing pass on good ideas, not a substitute for having them.

Pair this with our walkthrough on two-stage prompt systems (see the recent piece on handling refund emails) and you've got the editorial half of the workflow. The prompt drafts. The system constrains. Your taste decides.