How to Write a Video Ad Script With AI (2026)

To write a video ad script with AI, stop writing a screenplay and start writing a shot sheet: a numbered list where each line pairs one spoken sentence with the visual an AI tool needs to generate it, plus a duration. Lead with a hook in the first 1.5 seconds, keep the whole thing under 100 words, then feed each shot to your video generator as its own prompt.

The script is the prompt now

Every AI video ad dies at the same line: the one the generator actually has to read.

AI can generate video from a sentence. The hard part is writing a sentence it can actually execute, shot by shot.

That one shift breaks every video-ad-scripting habit you have. A traditional ad script is written for a human crew: a director reads it, interprets it, and fills the gaps with taste.

An AI video tool has no taste and fills no gaps. It renders exactly what the words describe and nothing they leave out.

So the script format changes completely. You are no longer writing dialogue for a person to perform. You are writing a set of generation instructions that happen to include dialogue.

Shots, not scenes. Timing, not stage directions. Prompts, not pages.

Here is the difference in the space of one beat. First, the old way, a screenplay line:

INT. KITCHEN - DAY. SARAH, 30s, frustrated, scrolls her phone at the counter. She sighs.

Feed that to an AI video generator and you get a coin toss: wrong age, wrong kitchen, a sigh you cannot hear, no camera framing, no length. It reads fine to a director and renders badly to a model.

Now the same beat written as a shot sheet line, the way an AI tool can execute it:

SHOT 1 (0.0-2.0s) | Visual: medium close-up, handheld, a woman in her early thirties leaning on a kitchen counter, morning light from a window behind her, scrolling her phone with a tired expression, shot on a phone camera, vertical 9:16. | VO: "i've made forty-seven ads this month and hated every single one."

Same idea, completely different artifact. The second one carries framing, subject, lighting, format, duration and the exact line, so the model has nothing to guess.

This guide teaches you to write the second kind on purpose. It sits inside our pillar on AI video ads in 2026, and it is the step that comes before making the video ad itself.

The bones do not change: hook, problem, solution, CTA

The format changes. The structure does not. Almost every video ad that converts moves through four beats, and a good one does it in under 30 seconds.

• Hook (0 to 2s): stop the scroll. A specific number, a mild warning, or a sentence that starts mid-thought.
• Problem (2 to 8s): name the pain in the viewer's own words. One problem, not three.
• Solution (8 to 22s): show the product doing the one thing that fixes it. Show, do not list features.
• Call to action (22 to 30s): one instruction. Not "learn more, sign up, and follow." One.

The reason to keep this in your head while you script for AI is simple: each beat becomes one or two shots on your sheet, and the timing above becomes the duration on each line.

The structure is the skeleton. The shot sheet is how you hang generation instructions on it.

Write the hook before you write anything else

The first 1.5 seconds decide whether the other 28 exist. So the hook is not the first line of your script, it is a separate project you finish first.

Write ten hooks before you write a single body line, then make them fight for the slot. Three patterns that reliably win:

• The specific number: "i tested 12 ad tools and only one made money." Odd, precise numbers read as real, not rounded marketing.
• The mid-thought open: "...so i finally stopped paying a agency $3k a month." Starting mid-sentence kills the ad-intro reflex that moves thumbs.
• The mild warning: "don't run this ad until you fix your first line." A small, specific caution beats a big vague promise almost every time.

Notice what none of them do: none say "introducing," none open on a logo, none name the product in the first line.

This is the one job worth handing to a language model first. Give it your rough idea and ask for ten hooks in that style, then delete the nine that sound like a brand wrote them.

A prompt that works: "write 10 video ad hooks for [product] aimed at [audience]. each must start mid-thought, use a specific number or a mild warning, and never say the brand name in the first line. casual, lowercase, no marketing words like game-changing or revolutionary."

The shot sheet: a script an AI can actually execute

Here is the whole method. Instead of a page of dialogue and stage directions, you write a numbered list of shots. Each shot is a single line with three parts:

• Timing: the shot number and its in and out point, like SHOT 3 (8.0 to 12.0s).
• Visual: the generation prompt for that shot, framing, subject, action, setting, lighting, camera, format.
• VO or on-screen text: the exact words spoken or shown during that shot.

Why a list and not prose: modern AI video tools generate in clips of a few seconds, not one continuous take. A shot sheet already matches that unit. Each line is a self-contained prompt you paste straight into your generator, and the durations tell you how many seconds to render.

In a tool built for this, like Flowjam, each line of the sheet becomes its own generation card, so the document you plan with is the thing you produce from. The script and the prompts stop being two documents. They become one.

A full 20-second shot sheet for a SaaS ad looks like this:

SHOT 1 (0.0-2.0s) | Visual: handheld medium close-up, a founder in a hoodie at a messy desk, late evening, laptop glow on face, vertical 9:16, phone-camera look. | VO: "i tested 12 ad tools this quarter and only one made money."

SHOT 2 (2.0-8.0s) | Visual: screen recording of a cluttered analytics dashboard, cursor moving between broken charts, slight zoom. | VO: "the problem was never the tool, it was that every ad took me a full day to write and shoot."

SHOT 3 (8.0-16.0s) | Visual: clean product UI of a video tool, a script turning into a finished vertical ad, smooth screen capture, bright. | VO: "now i write a shot sheet, paste it in, and get five ad variants before lunch."

SHOT 4 (16.0-20.0s) | Visual: the founder again, relaxed, slight smile, same desk now tidy, holding phone showing the ad. | VO: "link's in bio, make your first one free." | On-screen text: "start free"

Four shots, four prompts, one 20-second ad, a hook that lands in the first two seconds and a single CTA at the end. That is the entire deliverable.

The visual direction syntax AI tools understand

The "Visual" half of each line is where scripts for AI live or die. A model needs concrete, camera-language description, not a novelist's mood.

Order the words the way tools parse best: shot type, then subject, then action, then setting, then lighting, then camera, then format.

• Shot type: extreme close-up, close-up, medium, wide, over-the-shoulder, POV. Name it every time.
• Subject: specific and physical. "a woman in her early thirties in a grey sweater," not "a customer."
• Action: one verb per shot. Generators blur multi-step action, so keep it to a single beat.
• Lighting: "soft morning window light," "harsh overhead office light," "warm golden hour." This does more for realism than any other word.
• Camera: "handheld, slight shake," "locked-off tripod," "slow push-in." "Handheld" alone kills the too-perfect AI look.
• Format: "vertical 9:16, phone-camera look" for social, "cinematic 16:9" for landing pages. State it on every shot.

Two words earn their place in almost every ad shot: "handheld" and "phone-camera look." Studio-clean footage is exactly what signals "an AI or a brand made this" and gets scrolled past. Roughness reads as real.

Timing: match your shots to the length that sells

Because each shot on your sheet carries a duration, length is a decision you make on purpose, not an accident of how much you wrote. Three lengths cover almost every video ad:

• 6 seconds: one hook, one shot, one line, a bumper. Great for retargeting and YouTube pre-roll. No room for a problem beat, so the hook carries everything.
• 15 seconds: three shots. Hook, one problem-solution beat, CTA. The default for most paid social.
• 30 seconds: four to six shots. Full hook, problem, solution, CTA with room for one proof shot. The ceiling before completion rate drops.

Keep total spoken words near 2 per second: about 12 words for a 6-second ad, 30 for 15 seconds, 60 for 30 seconds. Write past that and your voiceover races the visuals and loses.

Decide the length first, then let it tell you how many shots your sheet gets. It is easier to fill a fixed frame than to trim a runaway one.

The prompt that writes your first draft

Once you understand the shot sheet, a language model can draft one for you in seconds. The quality of the draft is decided entirely by the brief you give it.

A generic prompt like "write a video ad script" returns a generic screenplay. This one returns a shot sheet you can almost paste straight into a generator:

You are writing a shot sheet for an AI-generated video ad, not a screenplay.

Product: [what it is, one line]. Audience: [who, and what they currently do instead]. Goal: [the one action]. Length: 15 seconds. Format: vertical 9:16, handheld phone-camera look.

Output a numbered list of shots. Each line must have three parts: (1) timing as SHOT n (start-end in seconds), (2) Visual: a generation prompt ordered shot type, subject, action, setting, lighting, camera, format, and (3) VO: the exact spoken line.

Rules: hook must land in the first 1.5 seconds and start mid-thought or with a specific number. One verb per shot. Total spoken words under 30. One call to action only. Casual, lowercase. Never use the words game-changing, revolutionary, seamless, or "introducing."

Run it, then run it again with a different hook. Generating five shot sheets costs you a minute and gives you five ads to test, which matters more than any single perfect script.

When the generator ignores your shot sheet

It will happen. You write a clean line and the model hands back the wrong face, a static shot where you asked for a push-in, or a second person who was never in the prompt. This is the part every other guide skips, so here is the fix list, in the order to try it.

• It rendered two actions at once, badly: you asked for more than one verb. Split the shot in two. "She picks up the phone and smiles" becomes SHOT a (picks up) and SHOT b (smiles).
• The subject keeps changing between shots: models do not remember your character. Lock it with a reference image or the exact same physical description, word for word, on every shot that person appears in. Copy-paste the subject clause, do not paraphrase it.
• It ignored the camera move: put the motion first and keep it simple. "slow push-in on" at the start of the visual beats "...with a slow push-in" at the end, and "handheld" survives better than "dolly along a curved track."
• It looks too clean and fake: add "phone-camera look, slight motion blur, natural light" and drop any word like "cinematic" or "4k" that pushes it back toward gloss.
• The lip-sync is off on a talking shot: shorten the VO for that shot. A four-second clip holds sync far better than a nine-second monologue, which is another reason the shot sheet's short durations help.

The meta-move behind all of these: regenerate one shot, not the whole ad. Because your script is already broken into per-shot prompts, you can re-roll the single clip that failed and keep the four that worked, instead of gambling the entire video on one render.

That is the quiet superpower of scripting this way. A screenplay is all-or-nothing. A shot sheet fails one line at a time.

Adjust the sheet for where it will run

The method is the same everywhere, but three dials change by platform, and you set them on the sheet before you generate a frame.

• TikTok and Reels: vertical 9:16, hook in the first 1.5 seconds, captions baked on, a native handheld look. Sound-on, casual, no logo up front. See the full TikTok walkthrough.
• YouTube pre-roll: assume the first 5 seconds are skippable, so the hook and the brand payoff both need to land before the skip button matters. A 6-second bumper needs the whole message in one shot.
• Feed and in-stream (Meta, LinkedIn): many viewers watch muted, so your on-screen text column has to carry the message on its own. Write captions that make sense with the sound off.

Before you generate anything, run the sheet through a five-point check. It takes a minute and saves you a pile of wasted renders:

• Does the hook land, and is it the strongest line on the page?
• Is there exactly one verb and one framing per shot?
• Is the subject described identically on every shot they appear in?
• Do the durations add up to your target length, and the words to roughly two per second?
• Is there one call to action, not three?

If a line fails the check, fix it in text now. A bad prompt is cheap to fix and expensive to render.

Edit it like it will be spoken, because it will

The first draft is raw material, not a finished ad. Three edits turn a generic AI script into one worth generating:

• Read it aloud. A video ad script is heard, not read. Anything you stumble over, a viewer's ear rejects. Cut it.
• Cut every filler word. "In order to," "the fact that," "basically." Vertical attention is short and every wasted word costs you a viewer.
• Sharpen the hook one more time. It is worth more than the rest combined. If the first line is not the best line, swap it.

One reusable trick: keep a short voice document, five to ten phrases your brand actually uses, a few words it never uses, and the sentence rhythm that sounds like you. Paste it into the prompt and the model stops sounding like every other AI ad.

The model gives you speed. You give it taste. That division of labour is the whole skill.

The whole method, in five lines

Write the hook first and make it fight for the slot. Keep the four beats: hook, problem, solution, one CTA. Draft it as a shot sheet, not a screenplay, one line per shot with timing, visual and VO. Write the visuals in camera language the model can execute. Then edit it out loud and generate five versions, not one.

Do that and your script stops being a document you hand off and hope for. It becomes the exact instruction set your AI tool renders, which is the only kind of ad script that survives contact with a generator.

Your first move takes ten minutes: pick one product, write three hooks with the patterns above, and turn the best one into a four-line shot sheet. That sheet is your first ad.

Then paste it, shot by shot, into a generator built for this. Flowjam takes a shot sheet like the one above and turns each line into a clip, which is exactly why writing the script this way pays off: the document you plan with is the document you produce with.

For the next step, see how the sheet becomes a finished clip in our guide to making a video ad with AI, the platform-specific TikTok ad walkthrough, or the broader how to make AI videos pillar.

Frequently asked questions

How do you write a video ad script with AI?

Write it as a shot sheet, not a screenplay. List the ad shot by shot, and on each line pair a spoken sentence with the exact visual the AI tool needs to generate it, plus a duration. Lead with a hook that lands in the first 1.5 seconds, keep the whole thing under 100 words, then feed each shot to your video generator as its own prompt.

What is the difference between a normal script and an AI video script?

A normal script is written for a human crew that interprets it and fills gaps with taste. An AI video tool has no taste and fills no gaps, so it renders exactly what the words describe. That means an AI script has to carry the framing, subject, lighting, camera and format for every shot, because anything you leave out, the model guesses at.

How long should a video ad script be?

Match it to the ad length: about 12 spoken words for a 6-second ad, 30 words for 15 seconds, and 60 words for 30 seconds, roughly two words per second. Decide the length first, then let it tell you how many shots your sheet gets. Writing past that budget makes the voiceover race the visuals and lose.

Can ChatGPT or another AI write a good video ad script?

It can write a strong first draft in seconds, but only if you brief it well. A generic prompt returns a generic screenplay. Ask it for a shot sheet with timing, a visual prompt and the exact line per shot, ban marketing cliches, and give it a short voice document. Then edit the draft out loud, because the model gives you speed and you supply the taste.

Why do AI video generators ignore parts of my script?

Usually because the line asks for too much. Keep one verb and one framing per shot, describe the subject identically on every shot they appear in, put camera moves first, and shorten the voiceover on talking shots to hold lip-sync. Because the script is already split into per-shot prompts, you can regenerate just the shot that failed instead of the whole ad.