The Script Is Also the Shot List
This is the structural difference between a voiceover script for AI video and any other kind of writing. In traditional production you write the narration and then someone finds footage for it. In AI video you are generating every frame, which means every line you write is a generation request whether you meant it that way or not.
So write it that way from the start. Two columns, or two alternating lines. Left is what is said, right is what is on screen. If you cannot name the visual for a line, that line is abstract and either needs rewriting into something concrete or cutting.
| Voiceover line | What is on screen |
|---|---|
| Most people quit in the first month. | A single figure walking away down an empty corridor, slow push in |
| Not because it is hard. | Same corridor, held, no movement |
| Because nothing happens for six weeks. | Cut to a calendar page, light shifting across it as time passes |
That table is a script and a shot list at the same time, and writing it in that form makes it obvious when a line has no visual. Lines with no visual are where AI videos turn into stock footage montages.
Write for the Ear, Then Test It Out Loud
A synthetic voice does not improvise around awkward phrasing the way a human reader does. A human hits a clumsy sentence and instinctively reshapes the emphasis to make it land. A synthetic read delivers exactly what you wrote, including the parts that do not work.
- Read every line aloud at normal speaking pace. If you stumble, rewrite. Your stumble is a preview of the listener's confusion.
- Cut every subclause. If a sentence has a comma in the middle holding two ideas together, split it into two sentences.
- Replace written words with spoken words. Nobody says 'utilise', 'moreover', or 'in order to' out loud.
- Kill the numbers you do not need. Spoken numbers are hard to hold. One number in a line is fine, three is noise.
- Watch for words that are ambiguous when heard rather than read. Homophones and unusual proper nouns are where synthetic voices go wrong most reliably.
The Opening Five Seconds
There is no presenter here to buy you goodwill. No face, no charisma, no warmth to carry a slow start. The opening line does all the work alone, which means it has to be the most interesting sentence in the entire script.
- Open with the conclusion or the contradiction, not the context. The context can arrive in line four, once someone has decided to stay.
- Do not introduce the topic. Start inside it. 'Most AI videos fail in the first three seconds' works. 'Today we are looking at AI video retention' does not.
- Keep the first line under about twelve words. Long openings read as throat clearing.
- Make the first visual match the first line exactly. A mismatch in the opening is the fastest way to lose someone.
Rhythm - The Thing Most Scripts Are Missing
Monotone is not really about pitch. It is about pace. A script where every sentence is the same length produces a read that sounds flat even when the voice itself is good, because the listener's ear stops finding anything to hold onto.
The fix is to vary sentence length on purpose. Long sentence that sets up the situation and gives the listener some context to sit inside. Then a short one. Then a pause. Then the point.
- Write pauses as their own line in the script so they survive editing. A pause you only imagined is a pause that does not exist.
- Follow every long sentence with a short one. This alone fixes most flat scripts.
- Put your most important sentence immediately after a pause. That is where attention is highest.
- Read a full page and listen for whether it has a shape. If it sounds like a list, it will play like a list.
How Long Should It Be
Work backwards from the video length rather than writing until you are done. A comfortable spoken pace is roughly one hundred and forty to one hundred and sixty words per minute for narration, and slower than that when the content is dense or you are leaving real pauses.
| Target video length | Rough word count | Approximate shot count |
|---|---|---|
| 30 seconds | 70 to 80 words | 6 to 10 shots |
| 60 seconds | 140 to 160 words | 12 to 18 shots |
| 3 minutes | 420 to 480 words | 35 to 50 shots |
Those shot counts are the number people underestimate. A three minute AI video needing forty distinct visuals is why so many end up as ten clips stretched thin. Knowing the number before you write means you write to it.
Writing Lines That Are Actually Generatable
The last pass is a practical one. Some perfectly good sentences describe things that current AI video handles badly, and it is much cheaper to catch that in the script than after twenty generations.
- Anything requiring readable text on screen. Signs, documents, phone screens, and labels render as garbled. Rewrite the line so the meaning does not depend on text.
- Anything requiring a specific brand or product. It will come out as an approximation that looks wrong to anyone who knows the real thing.
- Anything requiring precise hand action. Hands manipulating small objects is still where generations break down most visibly.
- Anything requiring a specific real person. Unless you are locking a likeness deliberately, avoid lines that demand a recognisable individual.
- Anything requiring an exact count of things. If the line says three people, you will spend generations fighting to get three.
None of these mean you cannot make the point. They mean you make the point with a different image. 'The invoice said four thousand dollars' becomes a hand setting down a folded document while a face reacts, and the voiceover carries the number.
The Checklist Before You Generate
Run this before a single clip gets generated. It takes five minutes and it is the difference between one generation pass and four.
- Every line has a named visual beside it.
- You have read the whole thing out loud and rewritten every stumble.
- No sentence contains two ideas.
- The opening line is under twelve words and contains the conclusion or a contradiction.
- Sentence lengths vary, and pauses are written in as their own lines.
- The word count matches your target duration.
- No line depends on readable text, a specific brand, precise hand action, or an exact count.
Short, practical drops on income paths, prompts that work, packaging for views, and getting paid. No spam, unsubscribe anytime.
Frequently asked questions
How long should a voiceover script be for a 60 second AI video?
Around 140 to 160 words. Narration sits comfortably at roughly 140 to 160 words per minute, and slower if the content is dense or you are leaving genuine pauses. Write to the target length rather than writing until you are finished, because an overlong script forces you to rush the read.
Why does my AI voiceover sound flat even with a good voice?
Almost always sentence rhythm rather than the voice itself. If every sentence is a similar length, the read sounds monotone no matter how natural the voice is. Vary sentence length deliberately, follow long sentences with short ones, and write pauses in as their own lines so they survive into the recording.
Should the script include the visuals?
Yes. In AI video you are generating every frame, so every line is effectively a generation request. Write the script in two columns, voiceover on one side and the intended visual on the other. Any line you cannot name a visual for is too abstract and needs rewriting or cutting.
How do I write an opening line for an AI video?
Lead with the conclusion or a contradiction, keep it under about twelve words, and make the first visual match it exactly. There is no presenter to buy you goodwill, so skip the topic introduction entirely. Filler openings like 'in this video' are the most common reason viewers leave in the first three seconds.
What should I avoid writing into an AI video script?
Lines that depend on readable on-screen text, specific branded products, precise hand manipulation of small objects, recognisable real people, or an exact count of objects. Current AI video handles all of these badly. Rewrite so the voiceover carries that detail and the visual carries the emotion instead.
How many shots does a three minute AI video need?
Roughly 35 to 50 distinct visuals. That number surprises most people and it is why long AI videos often feel padded - the shots get held far too long to cover the gap. If the shot count is impractical for your time budget, shorten the video rather than stretching each clip.
Last reviewed by David on August 16, 2026
