Writing Video Ad Scripts with AI: A Workflow That Keeps the Voice Human

What winning scripts sound like, measured

Before writing a script with AI, it helps to know what the target sounds like. We took the transcripts of the 1,752 winning video ads on our board (October 9, 2026, all with $50k+ in estimated spend and 30+ days live) and measured the opening of each, roughly the first 100 words, which is where voice matters most.

And the people in them:

Now the words models reach for when asked to write ad copy without examples:

That gap is the whole problem with AI scripts in one picture. Real winners are short sentences, contractions, "you" and "I". Default model output is long sentences, brand voice and launch words. The workflow below closes the gap.

Step 1: collect three to five transcripts in your format

Don't mix formats. A UGC review, a founder story and a voiceover listicle have different rhythms, and a model given all three will average them into something that matches none.

Pick winners in the format you're about to shoot, ideally live 90+ days. If you use Claude with the Ad Radar MCP connector, this is one prompt:

search_ads with tags vformat "ugc-talking-head", active, sort by
score, limit 15. Then get_ad on the 4 with the most days live and
give me their full transcripts, unedited.

Without the connector, transcribe from the Meta Ad Library yourself. Either way, keep the transcripts verbatim, including the "like" and the half-sentences. That mess is the voice.

Step 2: have the model describe the voice, not copy it

Ask for a voice profile before any writing. This turns vague taste into rules the model can follow.

Here are 4 transcripts of winning ads in the same format.
Describe the shared voice as rules:
- Average sentence length and how often sentences are fragments.
- Who speaks, to whom, and how they refer to the product.
- Reactions and filler words used, with examples.
- How the product claim is phrased (stated, shown, or reacted to).
- What never appears.
Quote examples for every rule. Do not write a script yet.

The quotes keep it honest. If a rule has no quote behind it, delete the rule.

Step 3: build the skeleton in seconds

Plan the beats against time, not words. At a normal speaking pace of about 2.5 words a second, a 45-second script is around 110 words.

Target: 45-second UGC review for [product]. Using the beat order of
the reference transcripts, give me a skeleton:
beat | seconds | job | what is on screen | max words
Hook must land in the first 3 seconds. Product on screen by second 8.

Two rules matter most here. The product should be visible early, and every beat needs something on screen, not only words. For the opening seconds specifically, see the first 3 seconds of a video ad.

Step 4: draft inside hard limits

Now the model writes, with the voice profile and the skeleton both in the prompt.

Write the script for this skeleton in the voice profile above.
Hard limits:
- Median sentence 9 words or fewer. At least 4 fragments.
- Contractions everywhere they'd be spoken.
- Speaker is [persona], talking to "you". Never "we" or "our".
- No: introducing, revolutionary, innovative, discover, transform.
- Product facts allowed: [list]. Anything else: [NEEDS FACT].
- Mark 2 lines as [CREATOR'S OWN WORDS] with a note on what to say.

The last limit is the most useful one. A script that dictates every word gets read like a script. Two open lines let the creator react the way the people in the reference ads do.

Step 5: run an AI-tell pass

Paste the draft back with a checklist and ask for a line-by-line audit. A second model, or a fresh chat, does this better than the one that wrote it.

Audit this script line by line. Flag:
1. Sentences over 15 words.
2. Lines a real person wouldn't say out loud.
3. Brand voice ("we", "our", "introducing").
4. Lists of three adjectives or benefits.
5. A closing line that summarizes instead of asking for the click.
6. Any claim not in the facts list.
Rewrite only the flagged lines. Keep everything else.

Point 4 catches the most common tell. Models love three of everything ("soft, durable and stylish"). Spoken winners rarely list; they react to one thing at a time.

Step 6: read it out loud, then cut

Read the script aloud with a timer, or have the creator do a table read on a call. Anything you stumble over, cut or rewrite. Then cut a 30-second and a 15-second version from the same script, keeping the hook and the offer word for word, so the editor has variants without a new shoot.

Here's a reference for how loose a high-spend script can be:

A one-page version of the workflow

The same approach works for the rest of the ad. For openers, see AI hook generator prompts; for briefing the creator who will say the lines, see how to brief UGC creators; and for the structural swipe from one donor ad, see the AI swipe workflow.

Figures marked as estimated spend come from Ad Radar's model of engagement on public Meta Ad Library ads. They are estimates, labeled as such, and are best used to rank ads against each other.