AI Script Writing for Short-Form Video

An AI model writes a 30-second script in seconds, and by default it is competent, generic and slightly too long. That is an input problem. Ask for “a script about my product” and you get the average of every acceptable script. A tighter brief fixes it.

What a model is good and bad at#

Good at: volume, variations on a fixed structure, rephrasing one claim ten ways, hitting a word count, matching a tone, writing in other languages.

Bad at: knowing what is true about your product, knowing what customers complain about, knowing which claims you can defend, and taking a risk in the first line. Models drift to the safe middle, and short-form dies there.

So you supply the specific, true, risky thing. The model supplies the variations and the discipline.

The five slots#

Every working short-form script fills five slots.

1. The address. Who it is for, as a situation, not a demographic. “If a client has ever asked for the same invoice twice” addresses someone. “For small business owners” addresses nobody.
2. The cost. What that situation costs them: time, money, risk, embarrassment. One sentence.
3. The turn. What changed, or the alternative. The product may appear here, not before.
4. The mechanism. Why it works. Viewers discount a benefit. They do not discount an explanation.
5. The action. One action, not a location. “Link in bio” is a place. “Start the trial, it asks for a card at the end” is an action with an objection removed.

Skip a slot only on purpose. A script with no mechanism is a claim. A script with no address is a broadcast.

Word counts by duration#

Spoken delivery runs at about 2.5 words per second.

Put the target in the brief as a hard number. Models follow word counts well, and it removes the most common defect: a voice rushing because the script is too long for the slot.

The brief#

A good brief has seven parts. You write it once and update it when something changes.

1. What you sell, in the customer’s words. The sentence a happy customer uses to recommend it.
2. Who exactly. Narrow enough that some people are excluded.
3. Three claims you can defend. With the evidence and any caveat.
4. Five objections you hear. Verbatim, including the unflattering ones.
5. What you will not say. Claims you cannot support, competitors you will not name, tones you will not use.
6. Two samples of your voice. Real sentences from your site or emails.
7. The action. What every video asks for, in one line.

Much of this can be extracted instead of written. Give Looptake your website URL and it builds brand-aware videos from your site’s audience, tone and offers. Agencies can store the brief per client as a brand kit, so a fix made once applies to every later video.

The prompt is not the main lever#

What moves output quality, in order:

1. Whether the model has specific, true material from the brief.
2. Whether the structure is set (the five slots, the word count).
3. Whether you generate many and reject most, instead of polishing one.
4. The prompt wording.

Ask for ten scripts and keep one. Polishing a weak script makes it acceptable. Generating ten gives you a chance at a good one. In Looptake, scripts are free to generate; credits pay only for AI media. See pricing.

Write the hook first, separately#

Most people write the body, then an opening to introduce it. That produces the typical AI-script defect: an opening that summarises the video instead of earning the stop.

Ask for ten openings. Reject any that would be true of a competitor. Pick one, then ask for a body that serves it. A model asked for the whole script at once optimises for coherence, and the first second needs something else. How to write video hooks covers the patterns.

The six-pass edit#

Run these in order before anything renders.

1. Delete the first sentence. AI openings warm up. If sentence two works as an opener, use it.
2. Read it out loud. Every place you breathe wants a full stop. Every stumble is a stumble for the voice.
3. Hunt for words nobody says. “Leverage”, “solution”, “robust”, “streamline”. A sentence with several of these needs replacing, not editing.
4. Check every claim against the landing page. A video that promises what the page does not support gets clicks that do not convert.
5. Count the words. Cut to fit. Do not speed up the delivery.
6. Cut a polite last line. “Check it out, you won’t regret it” is two seconds of nothing.

Spoken and on-screen text are two scripts#

The voiceover is what the viewer hears. The on-screen text is what the muted viewer reads. Captioning the voiceover word for word is not the same as writing on-screen text.

On-screen text carries the claim in the fewest words, readable as a still. Spoken text can be conversational. Write the on-screen line first: it is shorter, harder, and decides whether the video works with the sound off. Looptake adds word-timed captions from the voiceover automatically, so your effort goes into the on-screen headline.

A worked example#

Scheduling software for barbershops.

Weak first draft: “Running a barbershop is hard. Between appointments, walk-ins and no-shows, it’s easy to lose track. Our app makes managing your schedule simple, so you can focus on what you do best.”

Nobody is addressed, and there is no mechanism and no action.

Five slots, 54 words: “If your last no-show cost you forty minutes and you still texted them to check they were okay, this is for you. Deposits fix it, and most booking tools bury the setting. Ours turns it on by default: card on file, charged only if they miss. Turn it on in settings, it takes a minute.”

Address, cost, turn, mechanism, action.

For structures to put in the brief, use short-form script templates. To turn a finished script into a video with visuals, voiceover and captions, use the AI video generator.