How to Create Marketing Images With AI: Prompts and Edits
A repeatable three-step routine for turning a plain idea into a usable marketing image, then keeping every image after it consistent with the first.
A hook has about one second to justify the second one. AI is good at generating twenty of them, which matters because you cannot tell which works until you post.
Everything on social competes for the same thing: the decision, made in about a second, to keep going. That decision is made on the hook, and the hook is the part most people write last, quickly, after spending an hour on the thing nobody will see.
Hooks are also the ideal thing to generate in volume, because you genuinely cannot tell which one works until it is posted. Twenty candidates and a habit of testing beats one you were confident about.
Four mechanisms account for most of it:
What no longer works: “You won’t believe…”, anything that sounds like a headline from 2014, and hooks that promise something the content does not deliver — which trains the platform to stop showing you to people.
I'm posting a carousel about the three mistakes people make when repotting plants. Give me twelve first-slide hooks. Constraints: under eight words, no questions, no emoji, and each must work as large text on an image. Mix the mechanisms — specificity, open loop, contradiction, recognition — and label which each one uses so I can learn what works for my audience.
Write the caption for that carousel. First line has to stand alone before the "more" cut. Then three short paragraphs that add something the slides don't, and a closing line that invites a specific comment rather than "let me know what you think."
The first two seconds are the whole game, and spoken hooks follow different rules from written ones.
Give me ten spoken opening lines for a 45-second video about why sourdough fails in a cold kitchen. Rules: under three seconds spoken, conversational — not written-sounding — and the first three words have to do the work. Include two that start mid-thought as if the viewer walked in on a conversation. Also tell me what should be happening on screen during each.
Turn this into a 45-second script with timestamps: hook at 0:00, the shift at 0:03, and a reason to reach the end. Mark where a visual change should happen. No sign-off, no "make sure to follow."
Here's what actually happened: [describe a specific work situation and what you learned]. Write it as a LinkedIn post. The first two lines have to earn the "see more" click. Rules: no humblebrag, no "I'm humbled to announce", no single-sentence-paragraph staircase, no fake vulnerability. Concrete, specific, and it should be obvious I'm describing something that really happened. Under 200 words.
The genre’s clichés are strong enough that naming them in the prompt is necessary — models have absorbed a great deal of very bad LinkedIn writing.
Turn this idea into eight standalone posts, each under 240 characters and each making a complete point on its own. Direct, no threads, no "a thread 🧵", no numbered lists. Two should be strong enough that someone would disagree with them.
Give me eight title options for a video about [topic], plus the thumbnail concept for each. Titles under 60 characters, no all-caps, no arrows, and nothing that overpromises what's in the video. Then write the first 15 seconds of script for the strongest one — it has to confirm the viewer is in the right place and give them a reason to stay.
Write six bio options for [platform]. I do [X] for [audience], and what makes me different is [Y]. Under 150 characters, no job-title lists, no emoji dividers, no "helping X do Y" construction. One should be funny, one should be plain, one should be slightly odd.
The Marketing Expert assistant for angle and positioning, the Creative Writer for scripts, Image Gen for thumbnails and carousel graphics, and AI Goals for the posting cadence — consistency beats brilliance on every platform, and cadence is what fails first. See also AI social media assistant for planning content in your own voice.
If you ask for a caption, yes. If you supply a specific situation, ban the genre clichés, and pick from twenty options, no.
Fifteen to twenty per post. The cost is a minute and the variance between the best and worst is large.
Platforms rank on engagement, not authorship. Generic content performs badly because it is generic, which is a different problem with the same symptom.
Overpromising. It wins the click and loses the watch time, and watch time is what the algorithm actually measures.
Nobody reliably picks the winning hook in advance, including people who do this professionally. Generate twenty, post three, and let the numbers tell you which mechanism your audience responds to — then generate the next twenty against that.
Try it in ChatUp
Run the prompts above against the model that suits the task, keep the useful context across chats, and pick it back up on any device.
Try for Free