How to write an AI story video script: hook template, retention loop and prompts

A good AI story video script opens with a specific date, person and place, sets up an outcome the viewer expects, and hints that it went differently. After that, every part keeps one question open and opens the next question before the last one closes. Below is the hook template I use, taken from a 1.2M-view military history video. After it come the retention loop I add to every script and the prompts I run from research to upload.

What an AI story video is

An AI story video is a long story, usually 30 to 60 minutes, narrated over AI-generated images. No face, no camera, no acting. A lot of the audience is older and watches on a TV in the evening, letting one story run and then the next. Once the system is set up, a full video takes me under 30 minutes to make. That only works because the script does the heavy lifting, so this article is mostly about the script.

The hook template

The first 30 seconds decide whether someone watches for 40 minutes or leaves after 8 seconds. I’ve tested tens of hook structures across military history, romance, true crime and historical empire channels. This one holds the most viewers through the first 2 minutes:

On [specific date], [specific person with a single defining detail] [dramatic action verb] [specific location with atmosphere]. [Their target/opponent] was [vulnerable/unprepared/unaware]. [Person] expected [obvious outcome]. What [they] [encountered/walked into/discovered] instead [changed/rewrote/shattered] [larger consequence].

Here it is filled in, from the opening of a military history video that passed 1.2 million views:

On January 17th, 1781, a British officer named Banastre Tarleton rode toward a small cattle grazing field in upstate South Carolina with over 1,000 of the best-trained soldiers in the British army. His target was a ragged force of American troops, half of them untrained militia, who had a reputation for running at the first sight of a bayonet charge. Tarleton expected an easy victory. What he rode into instead was one of the most perfectly executed tactical traps in the history of warfare.

1. The specific date

“On January 17th, 1781” tells the viewer this really happened. It flips them into documentary mode. “A long time ago, a British officer…” reads like fiction, and trust drops with it. The date costs you five words and buys credibility for the next 40 minutes.

2. A person with one defining detail

A name creates a character. The defining detail, “with over 1,000 of the best-trained soldiers in the British army”, gives that character power and tells the viewer the scale of the threat. Use one detail per character. A second one in the hook slows it down, so save it for the body.

3. The vulnerable target

“A ragged force of American troops, half of them untrained militia” establishes the underdog. The power imbalance tells the viewer who to root for without a line of instruction, and once they pick a side they’re committed to seeing the outcome.

4. The expected outcome

“Tarleton expected an easy victory” confirms what the viewer already predicts, so viewer and character share the same expectation. It also begs to be broken. The viewer now watches to see how it breaks.

5. The payoff tease

“What he rode into instead was one of the most perfectly executed tactical traps in the history of warfare.” The viewer knows a trap exists and that it worked. They don’t know what it looked like, how it was designed or why it succeeded. That gap is enough curiosity to carry them through the next 19 minutes.

The same hook in other niches

My working targets: a hook should hold 85% or more of viewers through the first 30 seconds, and if it drops below 70% the rest of the video rarely recovers. Those are numbers from my own channels, not thresholds YouTube publishes. You’ll find the retention curve for each video in YouTube Studio analytics.

Keep one question open: the retention loop

A viewer stays for as long as there’s a question they need answered. The moment you answer everything, they leave. So the job is to keep one question open the whole way through. My scripts do that with four moves:

1. The cold open. Danger inside the first 3 sentences: a date, a place, a character and something about to go wrong.
2. The open loop. One question the viewer needs answered, handed out a little at a time. In the Cowpens example: what was the trap, and how did it work? They can tell he’s going to lose. They stay to see how.
3. The broken prediction. Let the viewer predict the obvious outcome, then break it. They expect the militia to break and run, and it does run. But the running was the trap: the officer chases them straight into the regulars waiting behind them.
4. The next loop. Open the next question before the last one closes. Most scripts die here: they answer the question and leave the viewer with no reason to stay.

My scripts run in 10 parts (chapters). Every part ends on a new question, and parts 3, 5, 7 and 9 end on the biggest ones. I put this block in the JSON template the script prompt reads from:

"retention_loop": "Open the video with danger in the first 3 sentences. After that, every part opens with something at stake, keeps one question open that the viewer needs answered, breaks one prediction the viewer is making, and ends by opening the next question before the current one closes. Cliffhangers at parts 3, 5, 7 and 9."

Then I test every part of a finished script with four checks, and rewrite any part that misses one:

• Something at stake in the first paragraph?
• One question left open?
• One prediction broken?
• A new question opened before the part ends?

Vary the intensity

A 40-minute story can’t run at peak intensity the whole time. Flat scripts feel exhausting, and random spikes feel unpredictable. The scripts that hold viewers best alternate high and low on a steady rhythm. For a 20-minute script:

Topics change every video. The rhythm stays the same, because it lives in the template rather than in each script. Viewers who watch three or four stories in a row are coming back for the rhythm as much as the topic.

The prompts, from idea to upload

These are the prompts behind the channels in my community. I write them for Claude, but nearly all of them work with any frontier chat model. Square brackets are the parts you fill in.

Stage 1: niche research

I connect the model to a source of live YouTube channel data and run this. Without a data connection, run the same brief by hand: it’s the 10-channel validation check from myth 1 of the faceless YouTube myths.

Find 10 YouTube channels in the [NICHE] space launched in the last 12 months with average views above 1,000 per video. Return the channel name, subscriber count, total videos, average views per video, estimated RPM range, primary audience demographic, and the titles of the top 3 outlier videos for each channel. Sort by average views descending.

Stage 2: outlier title analysis

Run this right after the research prompt. The model looks for the pattern across the outlier titles, so the 15 title ideas come out of data rather than guesses. More ways to find ideas are in the YouTube video ideas guide.

Analyze the outlier video titles from the channels above. Identify 1) the most common emotional triggers used in high-performing titles, 2) the structural formulas repeated across outliers (e.g. [person] + [action] + [nation]), 3) which title lengths perform best, 4) whether broad titles outperform specific titles. Based on the patterns, generate 15 title ideas optimized for outlier performance in this niche.

Stage 3: the script

This is the core prompt. Its structure keeps the model on a retention arc instead of a flat walk through information, and the rules strip out filler, rhetorical transitions and generic openers before they appear. Adjust [LENGTH] and the minute ranges to your niche’s best length (the video length data shows what gets views), and check the word count against your target runtime with the words to minutes calculator.

You are a professional long-form narrative scriptwriter for YouTube AI story channels. Write a [LENGTH]-minute voiceover script about [TOPIC] using the following structure:

HOOK (first 30 seconds): open with a specific date, location, and character. Establish immediate tension or stakes. The viewer must know within 30 seconds why this story matters.

CONTEXT (minutes 1-5): establish the wider context, the stakes beyond the individual story, and introduce the key characters with ONE specific, memorable anecdote per character establishing personality and motivation.

RISING CONFLICT (minutes 5-15): build tension through escalating events. Each scene raises the stakes. Alternate between high-intensity and breathing room. Never maintain the same intensity for more than 3 minutes.

CLIMAX (minutes 15-18): deliver the payoff the hook promised. Use visceral, visual language. Make the viewer feel present in the moment.

RESOLUTION + LEGACY (minutes 18-20): connect the story's outcome to something bigger. Show why this event mattered beyond itself. End with a line worth remembering.

RULES: skip filler phrases entirely, drop the "welcome to the channel" opener, and don't use rhetorical questions as transitions. Every sentence earns the next one. Use specific numbers, dates, names, and locations throughout. The script must be factually accurate.

“The script must be factually accurate” is an instruction, not a guarantee. Chat models still get dates, names and numbers wrong, so check the facts in a history or true crime script before you record it.

Stage 4: scene images

Paste the full script above this prompt. The model reads the story and writes 20 scene prompts with enough visual direction for the image model. Describing recurring characters the same way in every prompt keeps them close enough; perfect consistency isn’t worth hours of work. Official sizes and prompting tips for thumbnails, banners and in-video stills are in AI images for YouTube, and if you want moving shots, compare tools in the AI video generators guide.

Based on the following script, generate 20 image generation prompts for [IMAGE MODEL] producing super-realistic images depicting key moments from the story. Each prompt should specify character age, physical appearance (clothing, hair, expression, posture), setting (location, time of day, weather, architecture), lighting (direction, color temperature, mood), and camera angle (wide shot, medium, close-up). Maintain visual consistency by describing recurring characters with the same physical details in every prompt.

Stage 5: thumbnail

Send this with a competitor’s thumbnail as a style reference. Leaving the top corners clear matters because you need room for the text that does the emotional selling. A lot of this audience watches on a TV, so the text has to read from across a room. Before you upload, see how it looks next to other videos in the YouTube thumbnail tester.

I need a thumbnail image prompt for a YouTube video titled '[VIDEO TITLE]' in the [NICHE] niche. The thumbnail should 1) feature [NUMBER] characters with specific, exaggerated facial expressions conveying [EMOTION], 2) be set in [SETTING] with [LIGHTING CONDITIONS], 3) leave the top corners relatively uncluttered for title text overlay, 4) use [COLOR TEMPERATURE] tones matching the mood. Style reference: hyperrealistic cinematic film still, 16:9 horizontal, [GENRE DESCRIPTOR]. Include 8k detail, photorealistic skin texture, volumetric light, and shallow depth of field.

Stage 6: channel description

Paste a competitor’s description below the prompt. I use the output as the channel description and as the default video description.

Rewrite the following YouTube channel description to sound professional, engaging, and niche-appropriate for a [NICHE] channel. Keep the tone [TONE: mysterious/authoritative/warm]. Include a brief hook explaining what the channel covers, a line on upload frequency, and a subtle call to subscribe. Under 150 words.

Stage 7: translation for a second-language channel

One script, two languages, two channels. The rules make the translation read like a native script instead of a machine translation.

You are a professional literary translator and localizer for long-form narrative video scripts. Translate the script below from English into [TARGET LANGUAGE]. Rules: 1) translate 1:1, faithful and complete. Do not summarize, shorten, expand, or omit anything. 2) preserve the original meaning, tone, emotional register, narrative voice, pacing, and rhythm. Translate idioms and expressions into natural equivalents. 3) localize character names and place names into natural forms for [TARGET LANGUAGE]. Keep names 100% consistent throughout. 4) preserve exact formatting and paragraph structure. 5) output ONLY the translated script, with no preamble or notes attached.

Before you upload: YouTube’s rules

Two YouTube rules apply to every AI story video. The first is monetization: YouTube doesn’t pay for content that looks mass-produced or repetitive, so every video needs its own story, not the same video with a new title. The details are on YouTube’s AI content policy. The second is disclosure. YouTube’s GenAI disclosure page says:

“We require creators to disclose AI-generated or altered content that seems realistic.”

Its list includes content that “Generates a realistic scene that didn’t actually occur.” Photorealistic images of a real historical event can fall into that, so read the examples on that page and answer the AI-use question in YouTube Studio honestly. The same page says disclosing alone won’t limit a video’s audience or its eligibility to earn money.

One more thing worth checking before you blame a script: if a new channel gets no views at all, the script is rarely the cause. See why a new faceless channel gets 0 views.