The scroll is brutal. A user makes a stop-or-keep-going decision in the first 1.5 seconds of seeing a post, and almost all of that decision rides on the opening line and the visual. If your opening line sounds like every other product post in the feed, the algorithm-trained eye has already moved on. AI-written copy fails this test more often than not, because the average chatbot reply is, by definition, the average post in the feed.
The marketers writing copy that does stop the scroll are not relying on a single clever prompt. They are running a short three-prompt chain that pulls voice, angle, and tension out of the model in sequence. Here is the chain, with the prompts you can adapt to your own product.
Why one-shot prompts produce scrollable copy
The default behaviour of a language model is to summarise, balance, and hedge. Those three behaviours are the enemy of scroll-stopping. A scroll-stopper picks one specific tension, opens with it cold, and pays it off in the next two lines. A summary-balanced-hedged post opens with “Discover our new…” and the user is already two posts down.
HubSpot’s State of Marketing 2025 reports that 79% of marketers say AI reduces time on manual tasks. The marketers who report quality lifts as well as time savings are not using a different model. They are using a different workflow, one that forces the model to skip the safe opening it would naturally produce.
Prompt one: extract your voice from your existing best posts
Before writing anything new, give the model a sample of what you already publish so it can mimic the rhythm and word choice. The mistake is asking for “your voice” without giving examples. The model invents a voice that sounds like the average brand on Instagram.
“Below are five of my highest-engagement social posts from the last 90 days, with their engagement metrics. Read all five carefully. Identify and write down: (1) average sentence length, (2) typical opening pattern, (3) three signature words or phrases that recur, (4) what is consistently absent (emojis? hashtags? exclamation points?), (5) the rhythm of how the posts end. Output a one-paragraph ‘voice card’ I can paste into future prompts. Do not invent characteristics that are not in the samples.
POSTS:
[PASTE FIVE POSTS WITH METRICS]”
Save the voice card somewhere you can paste from. It is good for about 90 days before drift sets in and you need to regenerate it from a fresh batch of posts.
Prompt two: write to a tension, not a topic
This is the prompt that does the actual work. The instruction set forces the model to pick a specific tension before it writes a single line of copy, which is the move that separates scrolling-past from scrolling-stopping.
“You are writing an Instagram caption for [BRAND] about [PRODUCT/TOPIC]. Voice card: [PASTE]. Audience: [SPECIFIC SEGMENT]. Before you write, do this: list five tensions or unspoken frustrations this audience has about this topic. Pick the most specific one (not the most universal). Now write three caption variants, each opening with a single concrete sentence under 12 words that names that tension. No question openers. No ‘Have you ever…’ No ‘Did you know…’. Each caption is 70 to 110 words. Each ends with one open question, not a CTA. Mark which tension each variant uses.”
You will get three captions that each chase a different angle on the same product. The variants will not all work. Usually one of the three is clearly the strongest, one is interesting but slightly off, and one is a miss. The hit rate is roughly the same as a creative team brainstorm, which is the right benchmark.
The reason this prompt outperforms the standard “write me a caption” prompt is the tension list. By forcing the model to enumerate five frustrations before writing, you stop it from defaulting to the most obvious one. The most obvious tension is also the most-written-about tension, which is the reason every post about that product looks the same.
Prompt three: the scroll-stop test
Before any caption goes into your scheduler, ask the model to be the user.
“You are a 32-year-old in the [TARGET AUDIENCE] segment scrolling Instagram on a Sunday afternoon. You see this post for the first time. Read only the first sentence. Would you stop or keep scrolling? If you would keep scrolling, what specifically about that first sentence felt familiar or skippable? If you would stop, what hooked you? Then read the rest. At what point would you exit if you exited? Be honest, this is for a real test. POST: [PASTE CAPTION]”
Run this against each of your three variants. The model will identify weaknesses you missed: the opening that is too clever to land, the second sentence that betrays the first, the closing question that feels obligatory. You are not getting truth here, you are getting a structured critique that catches the patterns of AI-written tells before a real audience sees them.
A quick worked example
Brief: a small-batch coffee roaster launching a new single-origin Ethiopian. Audience: weekend home-brewers who own a burr grinder. Standard one-shot prompt produces something like “Discover our new Ethiopian Yirgacheffe, with bright notes of citrus and jasmine. Available now.”
Run through the three-prompt chain and you can end up with something like: “The grocery store coffee in your cupboard is six months old. This bag was roasted on Tuesday. We ship Wednesday. You taste the difference on Thursday. 200g, $24, single origin Yirgacheffe, no shipping fee on the first order. What you brew on a Saturday morning is the most underrated upgrade you can make this month.”
The second version is not better because the AI is more talented. It is better because the prompt forced the model to pick freshness as a tension, name a specific timeline, and skip the brand-voice opening that would have read as filler.
The failure modes that flatten AI-written social copy
The first is voice drift. Without a voice card pasted into every prompt, the model defaults to a vaguely upbeat, vaguely conversational voice that reads as generic Instagram. The fix is the voice card and the discipline of pasting it in every time.
The second is the AI-tell. Em dashes, “It’s not just X, it’s Y,” “elevate,” “redefine,” “embrace,” and the soft uplift ending where every post finishes with a vague aspirational sentence. Add an explicit ban list to your second prompt. The model honours ban lists better than it honours stylistic instructions.
The third is the missing constraint. “Write a caption for our new product” generates filler. “Write a 70-word caption that opens with a 9-word concrete sentence about a specific Sunday morning” generates copy. The constraint is the prompt.
If you want tools that wrap this loop in a UI, Jasper and Copy.ai both have brand voice features that handle the voice card automatically once you have trained them on enough samples. They are convenient. They are not magic. The underlying chain is the same: extract voice, write to tension, stress-test, ship.
What actually changes when this clicks
The honest payoff is not that AI writes great copy. It is that you stop staring at a blank page on Monday morning. The mental work of marketing copy is choosing the angle, and the angle is the part you still own. Once you have given the model a voice and a tension to chase, the typing is fast, the variants are plentiful, and the scroll-stop test catches the misses before they reach a real feed. Steady output, less burnout on captions, and posts that read like a brand instead of a placeholder.