How to Prompt ChatGPT, Claude or Gemini for Scroll-Stopping Ad Hooks

Use ChatGPT when you want a large, messy pile of hooks to filter. Use Claude when the category is regulated or the hook must stay inside a claims list. Use Gemini when the hook must match assets you already have in Google Ads or a live product URL. Switching models mid-batch without changing the spec just gives you three flavors of the same announcement.

ChatGPT is the volume tool: twenty to forty candidates, lots of duds, occasional sharp objects. Claude is the constraint tool: fewer lines, better at obeying 'do not name a medical result' and 'ten words max.' Gemini is the context tool: paste a product URL or an existing headline set and ask for openings that do not contradict the PDP. None of them is 'better at hooks' in the abstract. The miss is almost always the brief. If you type 'give me viral hooks,' every model will reach for 'Wait, this changed everything' because that sentence is what viral looks like in training data.

Do not ensemble blindly. Generating the same prompt in all three models and keeping everything triples the edit pile without tripling the number of distinct ideas. Pick one model for the batch, then — if the category is sensitive — run the shortlist through Claude for a claims pass. That two-step is cheaper than arguing about which chatbot is the most creative. Creativity here is a filter you apply, not a personality the vendor sold you in a launch blog post, and it will not survive a mute test on a phone.

Your prompt must include the job ('stop a mute scroll in two seconds'), the length ('ten words or fewer'), the formulas you want labeled, the customer quotes to steal nouns from, and a ban list of model-default openings. Also state what the hook must not do: name the product too early, promise a result you cannot show, or set up a joke that needs a second line.

Paste four to eight voice-of-customer sentences and circle the concrete nouns: the stained shirt, the 6am alarm, the bottle that leaked in a bag. Tell the model those nouns are mandatory raw material. Then list the formulas with one-line definitions: question, confession, complaint, number, pattern interrupt, direct callout. Ask for a labeled batch so you can see when the model collapsed three formulas into one paraphrase. Unlabeled lists hide that collapse. You will think you have diversity when you have synonyms.

Ban the costume: 'Are you tired of,' 'Stop scrolling,' 'What if I told you,' 'Imagine,' 'POV:', 'Wait for it,' and any line that could run for a rival SKU with the product name swapped. Add category-specific bans too — health superlatives, income claims, fake urgency. State the mute test in the prompt: 'This line must make sense with the sound off if it appears as on-screen text.' Models write for readers. Hooks have to work as captions. If you do not say that, you will get wordplay that dies the moment the phone is silent.

Ask for twenty hooks, four per formula, each on its own line with the formula tag. Do not ask the model to pick the best. Models pick the most typical. Your job is volume with labels so you can throw most of them away in the next step without losing the formula mix. Three 'polished' hooks is how you ship the average of the training set and call it a test.

Temperature and sampling differ by product, but the operational rule is the same: one shot of twenty is better than five chats of 'try again' that slowly converge on the same confession. If the first batch is timid, tighten the spec (more nouns, harder bans) rather than pleading for 'more energy.' Energy is how you get all caps and fake shock. Specificity is how you get a line a stranger can picture. If twenty lines share a skeleton — 'I cannot believe I [verb] this [noun]' — keep one and regenerate the rest with that skeleton banned.

For Gemini, attach the product image or URL so number-hooks and callouts do not invent a size or a price. For Claude, attach the claims table and require a silent tag on any hook that leans on a proof row. For ChatGPT, consider a follow-up that says 'now rewrite each survivor with one more concrete noun and one fewer adjective.' That second pass is often where usable lines appear. Do not merge the passes into one mega-prompt on the first try; you will not see which instruction the model dropped.

Kill any hook over ten words, any hook you cannot say in one breath, and any hook that still works if you swap in a competitor. Keep a shortlist of six. Clever rhymes and platform slang fail this scoring more often than plain sentences that name a real object. Plain is not the enemy of performance. Interchangeable is, because the feed already contains that line in someone else's voice.

Say each survivor out loud while holding your phone as if you were the avatar. If you smile at the wording but stumble on the mouth, it is not a hook — it is a tweet. Count words, not syllables of intention. 'I was today years old when I learned this stain rule' might be fine; 'I was today years old when I learned our proprietary enzyme blend exists' is a brochure that stole a meme. The competitor-swap test is the one teams skip because it feels mean. Do it. If the line is equally good for the next brand in the ad account, it will not stop a scroll that has seen twenty ads this hour.

Score the first frame at the same time, even as a note: what is on camera when the line is spoken. A complaint hook about a leaky bottle needs the leak, or the hands, not a smiling talking head in a clean kitchen. Hooks that cannot be pictured in one shot are usually essays. Mark those as script body, not as openings. You are not grading writing quality. You are grading whether a mute thumb has a reason to halt. If you cannot describe the frame in five words, the hook is not ready, no matter how much you like it in the chat window.

For the six survivors, write one line of visual direction: who is on camera, what they hold, and whether the first frame is a face, a mess, or a product action. Then drop any hook that has no honest frame you can actually film or generate this week. A scroll-stop is audiovisual even when the prompt only asked for words, and the model will not warn you that the shot does not exist.

This is where AI-written hooks fail in production. The model gave you a pattern interrupt that starts mid-action, and your avatar pipeline starts on a static head-and-shoulders shot. Either change the opening shot or change the hook. Do not hope the viewer will wait for B-roll. The first frame and the first line have to be the same idea. If you cannot shoot the mess, do not write the mess. Pick a confession or a question that a face can carry. Honesty about your shot list will save you from a week of 'the hooks tested poorly' when you actually tested mismatched audio.

If you generate the video with an AI avatar, put the first-frame note in the production brief next to the hook: 'start on hands twisting the cap, then cut to face on word four.' If you film a creator, send the six hook-plus-frame pairs and let them pick the three that fit their space. Creators will tell you which lines they cannot physically start on. Believe them. A slightly weaker line with a true opening shot will beat a brilliant line that begins after the viewer is already gone from the feed.

Pick three formula-different survivors. Keep the problem, demo, and CTA beats identical. Launch them as three ads, not three new campaigns, with enough budget to read thumb-stop or hold rate, then CPA. If you change the body at the same time, you will not know whether the hook did anything, and you will blame the model for a test you made unreadable.

Isolate the variable the same way you would in any creative test. Same avatar, same offer, same landing page. Name the files with the formula so reporting does not depend on memory. At 48 to 72 hours, kill the loser on hook rate or early drop-off; do not wait for a full ROAS story on a line people never heard. Promote a replacement from the unused three on the shortlist rather than prompting the model again from scratch. The unused shortlist is why you generated twenty. Using the chat like a slot machine after every losing ad is how the spec rots.

Klip Kanvas can render the three hook variants against one product brief and two avatars in a single sitting, which is the production shape this test wants: identical bodies, swapped openings, same captions except the first line. Watch each file once on mute before it goes to the ad account. If the on-screen hook text does not match the spoken hook, you accidentally tested a caption, not a line. Fix that before spend, not in the retrospective when the sheet looks noisy and nobody can name the variable.

When a hook wins, paste it into the spec as a positive example with its formula tag, word count, and why it passed the competitor-swap test. Add one negative example too: a line the model loves that you keep killing. The next batch will tighten. A spec that never eats its own results stays theoretical, and you will keep prompting from zero every Monday.

Few-shot beats adjectives. 'More like hook 14, less like hook 3' is an instruction a model can follow. 'Be punchier' is not. Keep a running list of five winners and five bans at the top of the prompt. Rotate the winners when the account fatigues so you are not cloning the same confession for six months. If performance drops, check whether the examples have become a new costume — every hook now starts with a number because a number won in May. Diversity of formula is a testing need, not a style preference. Put that need in the spec as a quota: at least three formulas in every batch of twenty.