rollTwenty seconds of speech is about forty words. That is the entire budget for a product video script, and almost every first draft blows it in the opening sentence.
The fix is to treat the script as an allocation problem. Every word has to be bought from somewhere, so decide up front what each section gets.
Write it out loud, first
Not after the visuals. Not alongside them. First.
Speech is the only element of a short video with a fixed, non-negotiable rate: about 2.2 words per second at a natural pace, slower if you want anything to land. Pictures can stretch and compress to fit, and audio cannot. If you cut visuals first and write narration into them, you get a script that either races or leaves dead air, and both read as amateur.
So: open a notes app, say the pitch out loud, and write down what you actually said. Your spoken version is almost always better than your written one, because speaking forces you to drop the qualifiers.
The five lines that waste your budget
These come up in nearly every draft, and none survive contact with a real viewer:
- "Introducing…" Two words, zero information. The viewer can see you're introducing something.
- "In today's fast-paced world…" Any sentence about the world rather than the product.
- "We're excited to announce…" Your excitement is not the value proposition.
- The full feature list. Five features at four words each is your entire budget spent on a menu nobody ordered from.
- The company origin story. Compelling at minute three of a podcast, fatal at second two of a reel.
Cut all five and most drafts drop from seventy words to forty without losing a single idea.
Write for the ear, not the page
A script is read aloud, so it obeys spoken rules rather than written ones:
- Short sentences. A comma is a breath; a period is a beat. Long sentences make the voice sound rushed even at normal speed.
- Numbers as words. "Ten dollars," not "$10." "Number one," not "#1." Whoever or whatever reads your script will otherwise stumble, and a stumble at second twelve costs you the viewer.
- One idea per sentence. If a sentence needs "and" twice, it is two sentences.
- End on a stressed syllable. "Roll it." lands. "You can try it out if you'd like" dribbles.
Silence is part of the script
The most common mistake in AI-narrated video is wall-to-wall talking. A script that fills every second of a twenty-second video is not a script, it is a monologue with pictures.
Leave gaps deliberately. After the hook, let one beat pass with only the product on screen. Before the final call to action, let the music breathe for half a second. Those silences are where the viewer catches up, and they make the words on either side feel more considered.
When we direct a reel, an unnarrated scene is treated as an intentional beat and never gets a caption read into it. Mixing a written performance with auto-generated filler is what makes narration sound like disconnected captions instead of one person talking.
Land the emphasis on the right word
Every sentence has one word doing the work. Find it, and make the picture agree with it.
"It reads your site and directs the whole thing in about a minute."
That sentence is about a minute. So the number appears on screen as the word is spoken, the sound design accents there, and the motion peaks there. Miss by 200 milliseconds and the whole thing feels automated, even though a viewer could never tell you why. This is the single highest-leverage detail in short-form narration, and it's covered in more depth in Anatomy of a 20-second product reel.
A worked example
Here is forty-one words for a fictional uptime monitor, allocated the way the diagram above suggests:
"Your status page said everything was fine. Your customers knew otherwise. Pulse checks every endpoint from twelve regions and tells you before they do. Ninety-second setup. No agent to install. Start free at pulse dot dev."
Six words of hook that name a real, specific failure. Nine on the problem. Fourteen on what it does. Six of proof, and it's checkable. Five to close, with the domain spoken as a human would say it.
Read it aloud with a stopwatch. It runs about nineteen seconds, which is exactly the length of the video it belongs to.
If you would rather not write it at all, that's the default path in Rollsee: the director writes the spoken arc from claims your own site already makes, distributes it across scenes, and renders it as one continuous performance. But the rules above are what it is following, and they work just as well in a text editor.