AI scriptwriting is using a tool like ChatGPT or Claude to draft short-form and ad video scripts fast, built around a hook that stops the scroll in the first second. Done right, you feed the model a clear brief and a hook formula, then direct it line by line. Done wrong, you get generic narration nobody watches. Here is how to do it right.
This is a different job from writing a long VSL. A VSL has minutes to build a case. A Reel, a TikTok, a YouTube Short, or a paid ad has about one second before the thumb moves. So the craft is compression: one idea, one hook, one payoff, spoken the way a human actually talks.
It is writing the spoken words and on-screen beats of a video with an AI model as your drafting partner, then editing for voice and pace. You are not asking the model to "make a video." You are asking it to write the words a person says and the text that flashes on screen, in the rhythm of short-form.
The reason it works is speed of iteration. You can generate ten hook variations in the time it takes to write one by hand, then keep the one that actually lands. The reason it fails is that most people stop at the first draft. The model writes a competent, slightly robotic script, they post it, and it dies. The win is in the direction you give and the editing you do after.
Think of three script types you will write most: a short-form organic script (Reel, Short, TikTok), a paid ad script (a video built to sell, with a clear call to action), and a talking-head explainer. All three share the same spine: hook, hold, payoff, ask. The mix just changes.
Definition: hook-first scripting. Writing the first line before anything else, because the first one to three seconds decide whether the rest of the script gets watched at all. Everything downstream serves the hook, not the other way around.
Because retention is the whole game, and retention is won or lost in the opening seconds. If people swipe past the first line, nothing else you wrote matters. The platform also reads that early drop-off and stops showing the video, so a weak hook caps your reach twice.
A hook does one of a few jobs fast: it names a specific person ("If you run Facebook ads and your cost per lead just doubled…"), it makes a claim that creates tension ("Most short-form advice is keeping your videos small"), it opens a loop the viewer needs closed ("I deleted my three best-performing posts on purpose. Here is why"), or it shows a result and withholds the method. The common thread is a reason to keep watching that lands before the brain decides to scroll.
Here is the practical move: write the hook separately from the script, and write a lot of them. Ten hooks, pick one. If you want the deeper mechanics of why curiosity gaps hold attention, open loops in copywriting is the companion read. For the writing of the hook line itself, how to write a headline that converts carries straight over, since a scroll-stopping first line is a spoken headline.
You are a short-form video hook writer. I make videos about [TOPIC] for [SPECIFIC AUDIENCE, e.g. solo founders running their first paid ads]. Write 10 opening lines (hooks) for a video about [ONE SPECIFIC IDEA]. Rules: - Each hook is one spoken sentence, under 12 words. - Mix these angles: call out the specific viewer, make a counterintuitive claim, open a curiosity loop, show a result and withhold the how. - No "in this video," no "let me tell you," no throat-clearing. - Written to be said out loud, not read. After the 10, tell me which 3 you'd test first and why, in one line each.
You give it the hook you already chose, the one idea, the audience, the length, and the structure you want it to follow. The more you decide before prompting, the less generic the output. A vague brief gets you a vague script every time.
The structure that holds short-form is simple: hook, then a hold beat that pays off the curiosity you opened, then the meat delivered as one tight point (not three), then a single clear ask. Tell the model to write for the ear, with short lines, contractions, and natural pauses. Ask it to mark on-screen text separately from spoken lines so your editor knows what to put on screen.
Say you sell a budgeting template and your idea is "people overspend on subscriptions they forgot about." The hook might be "You are probably paying for three apps you stopped using months ago." The hold beat names the feeling. The meat is the one move: open your phone settings, read the subscription list out loud. The ask is soft: follow for the rest. One idea, start to finish, under 45 seconds.
Write a [30 to 45 second] short-form video script. Topic: [ONE IDEA] Audience: [WHO] Chosen hook (use this exact line first): [PASTE YOUR HOOK] Goal of the video: [follow / save / click link] Structure: 1. HOOK (the line above) 2. HOLD: one sentence that pays off the curiosity the hook opened 3. POINT: deliver ONE useful idea, no more, in plain spoken language 4. ASK: one specific call to action Format the output in two columns: - SPOKEN (what I say to camera) - ON-SCREEN (the text overlay for that beat) Voice rules: short sentences, contractions, no jargon, no em dashes, write the way a real person talks. Keep total spoken words under [110].
A note that saves you grief: tell it to keep the script under a word count, because spoken pace runs roughly two to three words per second. A 45-second video is around 110 to 130 spoken words. If the model writes 250, no editing will make it fit.
An ad script carries a sharper job: it has to sell, so it leans harder on the problem, the proof, and the offer, and it always ends with one unmissable call to action. Organic content can earn a follow with pure value. An ad has to move someone toward a click or a purchase inside the same 30 to 60 seconds.
The structure shifts from "one idea plus a soft ask" to a persuasion arc: hook, problem you agitate just enough to be felt, the turn (your product as the answer), proof or specifics that make it believable, and a direct call to action. The hook still rules, but now the back half has to close. If you are running these on platforms, pair the script with the targeting and creative-testing thinking in how to use AI for ads, and treat each hook as its own test inside an ad creative testing framework.
One discipline that separates ad scripts that sell from ones that just describe: handle the obvious objection inside the script. If the price is the wall, name it and reframe it. If "I have tried this before" is the wall, acknowledge it. Objection handling in copy translates directly to spoken ad scripts, because the camera is just copy with a face.
Write a [45 to 60 second] direct-response video ad script for [PRODUCT]. Audience: [WHO, and the pain that brings them] Core promise: [THE ONE OUTCOME the product delivers] Price / offer: [OFFER DETAILS] Biggest objection to overcome: [OBJECTION] Call to action: [WHAT they click and where it goes] Structure: 1. HOOK: stop the scroll, name the viewer or the pain (under 12 words) 2. PROBLEM: agitate the pain in 1 to 2 lines, no piling on 3. TURN: introduce [PRODUCT] as the answer 4. PROOF: one believable specific (a use case or a plain mechanism, do NOT invent stats or fake testimonials) 5. OBJECTION: name the objection above and reframe it in one line 6. CTA: one clear action, said with confidence Two columns: SPOKEN and ON-SCREEN. Conversational, no em dashes, under [140] spoken words. Give me 2 versions with different hooks.
You edit for your own voice, you cut the filler the model loves, and you read it out loud before you ever record. A first draft is raw material, not a finished script. The model gives you structure and speed; you give it a human sound.
The tells are predictable. Hollow openers ("in today's digital landscape"), perfectly balanced three-item lists in every paragraph, words like "delve" and "unlock," and a smooth, featureless rhythm where every sentence is the same length. The fix is to break the pattern: cut the first line if it is throat-clearing, vary sentence length hard (a four-word line next to a long one), swap any word you would not say to a friend, and add one specific detail only you would know. Reading aloud catches the rest, because anything that trips your tongue will trip the viewer's ear.
You can also make the model self-edit. After it drafts, ask it to flag its own AI tells and rewrite them. For a fuller system on this, write copy with AI without sounding like AI goes deep, and if you are building a repeatable content motion, how to create content with AI sets up the wider workflow this script slots into.
Here is a video script you drafted: [PASTE SCRIPT] Now act as a tough editor. Do three passes: 1. Mark every AI tell: hollow openers, "delve/unlock/supercharge" type words, repetitive three-item lists, and any sentence that sounds written instead of spoken. 2. Rewrite the script to fix them: vary sentence length sharply, use contractions, cut filler, keep one specific concrete detail. 3. Read it for pace and cut any line that does not earn its place. Match this voice sample of mine: [PASTE 3 to 4 SENTENCES YOU WROTE]. Return only the final cleaned script.
It runs in five steps: pick one idea, write and choose a hook, draft the script with structure, edit it to your voice out loud, then build a repeatable batch system. The first four get you one good video. The fifth gets you a content engine.
Start with the idea, singular. One video, one idea. Then write ten hooks and pick the strongest. Then draft the script with the structure prompt, capped at a word count. Then edit aloud and cut. Once that loop feels natural, batch it: give the model a list of ten ideas and have it produce ten hooks each, then script the winners in one sitting. That is how creators ship daily without burning out. If you want this to feed a real machine across formats, how to repurpose content with AI shows how one script becomes a Reel, a Short, and an ad.
The mistake to avoid: scaling before the loop works. If your hooks are weak, batching just produces more videos nobody watches. Get one video to hold attention, then turn up volume.
The best tool is the one you will iterate in fast. ChatGPT and Claude both write strong short-form scripts when you give a clear brief and a chosen hook. The model matters far less than your prompt and your editing. If you want a broader rundown of options for the writing side, the best AI copywriting tools breakdown covers the field.
Match the words to the runtime, since spoken pace is roughly two to three words per second. A 30-second video is around 75 to 90 spoken words, and a 45-second one is around 110 to 130. Always cap the model with a word count in your prompt, or it will overwrite and the script will not fit the runtime.
Let AI write ten hook options, then you pick and sharpen. The model is great at volume and angle variety, which is exactly what hook writing needs. Your job is taste: choosing the line that actually creates tension for your specific viewer, then tightening it to sound like you said it.
No, platforms rank watch time and retention, not who wrote the words. A well-edited AI-assisted script that holds attention will outperform a hand-written one that bores people. The risk is not the AI, it is shipping the unedited first draft. Edit for voice and the platform ranks the watch time, not the byline.
Use a persuasion arc for ads and a single-idea structure for organic. Organic content earns a follow with one useful point and a soft ask. An ad has to hook, agitate the problem, present your product, handle one objection, and drive a clear click, all inside a minute. Same hook discipline, different back half.
Want the prompt library and the hook formulas in one place, plus the rest of the AI marketing playbook as it ships? Join the free Asset Academy email list and we will send the working templates straight to your inbox, no fluff, just the stuff you will actually use this week.
Inside the Asset Academy community we build the copy, funnels, and offers together, with the prompts and the feedback. $96/mo, or save with annual.
Join the community →The community where we build the copy, funnels, and offers together, with the prompts and live feedback.
Join the Community →