AI Video Script Generation

Start with format and seconds, not the topic

The most common reason AI video scripts are unusable is length. A model asked to "write a script about our new service" produces an article with line breaks: too long, too dense, and written for the eye rather than the ear. Decide two things before prompting:

1. • The format: short social video, explainer, long-form YouTube, or ad.
2. • The target length in seconds, which sets the word budget.

Set the word budget from your own speaking rate

Speaking rates vary a lot between people and between delivery styles, so measure yours rather than using someone else's figure. Read 100 words of a script aloud at the pace you would use on camera and time it.

Worked calculation (illustrative): if 100 words take 40 seconds, your rate is 100 ÷ 40 = 2.5 words per second. A 45-second video therefore has a budget of 45 × 2.5 = about 112 words, minus a few seconds for pauses and on-screen moments with no speech. Aim for around 100 words and give that number to the model.

Structures by format

Script structures by video format

Format	Typical length	Structure	Common AI failure
Short social video	15 to 60 seconds	Hook in the first line; one point; one proof or example; one action	Tries to make three points and makes none
Explainer	60 to 120 seconds	The problem in the viewer's words; how it works; what it looks like in practice; next step	Describes features instead of showing the mechanism
Long-form YouTube	5 to 20 minutes	Promise of what the viewer will learn; chaptered sections; recap only if the video is long; next video or action	Padding between sections; weak chapter breaks
Video ad	Platform dependent	Hook; problem or desire; offer; proof; action, with the brand visible early	Claims the business cannot substantiate

For long-form YouTube, plan chapters in the script. YouTube Help says chapters come from timestamps in the description: the first must be 00:00, there must be at least three in ascending order, and each chapter must be at least 10 seconds. Asking the model to mark chapter breaks with titles makes the description easy to build afterwards.

For LinkedIn video ads, LinkedIn's specifications allow videos from 3 seconds to 30 minutes, so length is a creative decision there, not a platform limit.

The two-column script template

Video scripts need to say what is heard and what is seen, side by side. Ask for this format every time:

Two-column script template

Time	Audio (what is said)	Visual (what is on screen)	On-screen text
0:00 to 0:03	Hook line	Opening shot	Short caption of the hook
0:03 to 0:20	Main point	Supporting footage, screen recording or presenter	Key number or phrase
0:20 to 0:35	Proof or example	The example being shown	Source or label if a figure is shown
0:35 to 0:45	Action	Presenter or end card	The action and where to do it

A prompt that produces usable scripts

Write a [format] video script, [N] seconds long, with a maximum of [W] spoken words. Audience: [who, and what they already know]. The one point: [point]. Proof we can use: [paste only real, checkable proof]. Action: [one action]. Use this structure: [paste structure from the table]. Present it as a table with columns Time, Audio, Visual, On-screen text. Write for speaking: short sentences, no parentheses, no lists read aloud, numbers written as they are said. Give three alternative hook lines at the end.

The "written for speaking" rules matter. Text that reads well can sound stilted aloud: long sentences, brackets and bulleted lists are hard to say naturally.

Illustrative short script

Illustrative example for a fictional bike repair shop; not a client. Target: 30 seconds. At 2.5 words per second the ceiling is 75 words; this script uses about 50, leaving room for pauses and the product shots.

• • 0:00 to 0:05. Audio: "Your gears slip on hills? It is probably not the gears." Visual: close-up of a chain slipping. On-screen: "Gears slipping?"
• • 0:05 to 0:20. Audio: "Nine times out of ten, a worn chain stretches and stops sitting on the teeth properly. You can check it in ten seconds with a chain checker." Visual: mechanic using a checker. On-screen: "Check the chain first."
• • 0:20 to 0:30. Audio: "Bring it in and we will check it free while you wait." Visual: shop front. On-screen: "Free chain check."

Note the "nine times out of ten" line. It is the kind of specific-sounding claim models add freely; in a real script it needs evidence from the shop's own records or must be softened to "often".

Checks before filming

1. • Read it aloud at speed. Anything you stumble on, rewrite.
2. • Check every claim. In UK advertising, objective claims need evidence held before publication. Remove figures you cannot support.
3. • Plan captions. W3C's accessibility guidance (WCAG success criterion 1.2.2) expects captions for prerecorded audio in video, including who is speaking and meaningful sounds. Captions also help viewers who watch without sound. Keep the audio column clean so it doubles as the caption source.
4. • Check disclosure rules if you use synthetic media. YouTube requires creators to disclose realistic content that is AI-generated or altered, such as making a real person appear to say something they did not, or generating a realistic scene that did not happen. A script written with AI and performed by a real person does not fall under that rule; an AI-generated presenter or voice that looks real may.
5. • Confirm the action is live. The URL, offer or booking link in the script must work on the day the video goes out.

For how video fits into a wider production workflow alongside other formats, see the content production bottleneck.

FAQ

Can AI write a script that sounds like me on camera?

Closer, if you give it transcripts of you speaking rather than your written work. Spoken and written voices differ; a transcript of a past video or call is the better example for scripts.

Should the hook be written first or last?

Last, or at least revisited last. Once the point and proof are fixed, the model can offer hooks that promise exactly what the video delivers, and you pick the strongest of several.

Do I need to tell viewers the script was written with AI?

YouTube's disclosure requirement concerns realistic AI-generated or altered visuals and audio, not scripts written with AI assistance and performed by real people. Check each platform's current rules if you use synthetic presenters or voices.

How many script variants should I generate?

One full script and three to five hook lines is usually enough. On short formats the opening line decides whether anyone keeps watching, so that is where variation pays off.