How to Use AI to Write YouTube Scripts: Hook First

Don't ask AI for a script. Ask it for a hook, and refuse the first one.

How to use AI to write YouTube scripts, in one sentence: paste your voice and a one line idea into one thread, then walk it through five asks in order. The promise, the outline, the hook, the draft, and a cold read from a different model. You hold the pen at every step.

The shortcut is the bare prompt. "Write me a ten minute script about sour espresso." What comes back is a script nobody wrote, and you can hear it by the fourth sentence.

So can your viewers. This summer, Hank Green apologized to his 3.2 million subscribers after viewers heard the line "I appreciate the pushback" in the middle of an educational video and guessed a chatbot had written it. TechCrunch's account says he told them the line was a reply to his guest, that he had used ChatGPT to research the script, and that his chatbot use was "not healthy." He said he would post less.

One stray line, and a creator with an enormous audience was explaining his process in public. You don't need a stray line to get caught. A script that sounds like nobody does its own damage, only slower.

Use AI to build the script in pieces you can inspect, and put your own judgment between every piece.

Here is the job from start to finish. A one line idea goes in, a ten minute script comes out, and the hook gets rewritten until it holds. It works like dialing in a shot of espresso: change one thing, taste it, then change the next. I ran it in a single Magai thread, so every prompt below is the one I typed, every reply is judged, and the two places it went wrong are in here.

How to use AI to write YouTube scripts in this order, one ask at a time

The order matters more than the model. Each ask hands the next one something to work from, and the thread remembers all of it. A single prompt can't do that, because it has nothing to remember.

Say you run a home espresso channel. You post weekly, you are forty videos in, and this week's idea is one line: why your home espresso tastes sour, and the one dial to turn first. That is the idea I used. It is thin on purpose, because most weeks start that thin.

The whole run fits on one screen.

A cleanup pass follows the five: the timing check, the cuts and a scan for anything the chatbot left behind.

You need two things before the first ask. The first is a sample of your own voice. Paste about 300 words from the transcript of your best video, auto captions included, because "dry and blunt" is a description and your captions are the thing itself.

The second is your speaking pace. Read a page of an old script out loud against a timer, then divide the words by the minutes. I used 150 words a minute for the example channel. Measure yours, because it decides every number in the timing step.

Start with the promise, not the script

A video is a promise with a runtime. If the promise is wrong, no hook and no outline repairs it, so the first ask has nothing to do with writing.

The first message in the thread, as typed:

I make a weekly YouTube channel about home espresso. I talk straight to camera, dry and blunt, short sentences. I say "pull a shot," never "game changer," and I never open with "welcome back to the channel." My viewers own a $400 to $700 machine and a decent grinder.

Idea for this week: why your home espresso tastes sour, and the one dial to turn first.

Before we write anything, give me three versions of the promise this video makes to a viewer. One sentence each: who it's for, what they can do after, and what they'd type into the search bar. Then tell me which one I can pay off in ten minutes.

I gave it adjectives because the example had to stay short. You should give it the transcript. The reply, in part:

1. For home baristas whose shots taste sharp and sour: you'll know the one thing to change first and stop guessing. Search: "why is my espresso sour."
2. For people about to buy a new grinder to fix sour shots: you'll know whether to spend $300 or turn a setting. Search: "espresso sour new grinder."
3. For beginners who want one routine that fixes sour, then bitter. Search: "how to dial in espresso."

Pay off in ten minutes: 1. Promise 2 needs side by side grinder footage you may not have. Promise 3 is two videos.

I took promise 1. Promise 2 is the better business, so I parked it as next week's video, which costs nothing because the thread keeps it. The reply also caught something I would have found out in the edit bay: promise 3 is two videos wearing one title.

It judged each promise by what the footage would have to look like, and that is the question a promise has to survive. It never needed to know my machine.

One more habit before you move on. Before you trust the thread with a paragraph, ask it to describe your voice back to you in three lines, then correct the description. It costs one message, and it saves you from rewriting the same bad habit in every beat.

Ask for the outline before a single sentence of script

An outline is where a ten minute video gets its shape, and a model is good at shape. It is bad at knowing which shape is yours, so you ask for the outline as beats with a clock on them.

The second message, as typed:

Go with promise 1. Outline the video as beats with a time budget that adds up to ten minutes. For each beat, give me one line on what the viewer learns and one line on the question it leaves open for the next beat. Don't write any script yet.

The reply, in part:

1. Hook, 0:00 to 0:45. Leaves open: why the grinder?
2. What sour means, 0:45 to 2:30. Leaves open: so what made the water hurry?
3. Turn the grind finer, 2:30 to 5:00. Leaves open: how far, and how do I know?
4. Why not touch anything else yet, 5:00 to 6:45. Leaves open: what does a fix look like?
5. The three shot check, 6:45 to 8:45. Leaves open: what if it still tastes sour?
6. When it's still sour, 8:45 to 10:00. Leaves open: the channeling video.

The beats add up to exactly ten minutes. Then I read the order and found the problem. Beat 3 tells the viewer to turn a dial before beat 4 explains why every other dial can wait. A stranger would hear that as an instruction with no reason.

So I swapped them: "Swap beats 3 and 4. Explain why the other settings wait before you tell them what to turn."

Look at the question on every line. Those are your re-hooks, the small lines that give a viewer a reason to stay for the next beat. They are the part of the outline I would never let it skip, because they are the first thing you will forget at beat four, when the draft is going well and you are tired.

Write the hook three ways, then kill two

The hook gets the most rewriting because it carries the most weight. YouTube's retention report has a number for it: the intro percentage, the share of your audience still watching after the first 30 seconds. The first sentence also has to confirm what the title promised, or the viewer assumes they clicked the wrong video and leaves.

So I never ask for a hook. I ask for three, each a different move, and then I argue with them. Round one, as typed:

Write the hook for beat 1, about forty seconds spoken. Give me three, each a different move: one opens on the viewer's problem, one opens on a belief most people hold that I'm about to take apart, one opens on a result they can see. No "welcome back." No "in this video." The first sentence of each has to confirm what the title promises: "Why Your Espresso Tastes Sour (Fix This First)".

What came back:

A, the problem: "Your espresso tastes like lemon water, and you've already tried three fixes. New beans. Hotter water. A longer shot. Here's the one dial you skipped, and it takes ten seconds to turn."

B, the belief: "Everyone says sour espresso means your water is too cold. It doesn't. Sour means the water left the coffee too soon, and the fix is on your grinder, not your boiler."

C, the result: "Same coffee, same beans, same machine. The cup on the left is sour enough to make you wince. The one on the right is sweet. I changed one thing, and it wasn't the temperature."

Hook B is the closest to a video, and it is wrong. Temperature does affect extraction, so "It doesn't" is a claim the video would have to walk back in beat 3. A hook that gets retracted by minute three costs you the viewer at minute three. This is the first thing that went wrong, and I only caught it because I read B against the outline instead of against my taste.

Hook C is the best idea and the worst draft of words, because it needs the side by side cups and I would have to shoot them. Hook A is a familiar opener for this topic. So B gets the rewrite, and I told it exactly what was wrong.

Round two, as typed:

Keep B's structure but fix the overclaim. Temperature does matter a little. The hook can say the grind is the first thing to change, not that temperature is irrelevant. Give me four versions of just the first two sentences, shorter, numbered.

The reply, in part:

1. Most advice says sour espresso means your water is too cold. Sometimes. It's rarely the first thing to check.
2. Sour espresso is a timing problem. The water left your coffee before it finished the job.
3. Before you touch the temperature, buy new beans, or blame your machine, change one thing: the grind setting.
4. Sour means under-extracted. Under-extracted usually means too coarse. Fix that first, and most sour shots are over by lunch.

Number 2 holds. It is true, you can check it with a cup, and it sets up beat 2 without promising anything the video can't pay. Number 3 is the second half of the same hook. I fused them. I cut number 4 for the reason hooks die: "most sour shots" is a number nobody has.

Round three is the one I would never skip. I ask the model to try to break the hook, as typed:

Here's the hook: "Sour espresso is a timing problem. The water left your coffee before it finished the job. Before you touch the temperature, or buy new beans, or blame your machine, change one thing: the grind setting."

Run the hold test. (1) Does the first sentence confirm the title? (2) Does it open exactly one question the video answers in the next minute? (3) Is every claim one I can pay off with no exceptions I'd have to walk back? (4) Does any of it sound like someone else's channel? One line each. Then name the weakest word.

1. Yes. "Sour espresso" is in sentence one.
2. Yes, but the last sentence names the answer, so the open question is why, and beat 2 pays that at 0:45.
3. Yes, with one watch: "timing problem" is true, but a viewer may hear shot time. Beat 2 has to say what it means in its first ten seconds.
4. "Finished the job" is a stock phrase.

Weakest word: "dial," in your title idea. Your viewer's machine has dials. Your grinder has a setting.

I kept "finished the job," because the sentence before it carries the claim and a little stock language is fine when it isn't doing the work. I acted on the other flags. Beat 2 now opens with "Sour is what unfinished coffee tastes like," which says what "timing" means in plain words, and the idea line stopped saying "dial." The hook that went into the draft is the one in round three, and it is 35 words, about 14 seconds of a 45 second beat.

Draft one beat at a time and make the model flag what it added

Never ask for the whole script in one go. A model handed 1,500 words to write tends to give you the average of every video about the topic. A model handed one beat, your notes and a short leash has nothing to average.

The prompt for beat 2, as typed:

Draft beat 2 in my voice, spoken, first person, short sentences. Use only my notes below. Anything you add from general knowledge, mark [CHECK] right after it. Don't use "game changer," "unlock," or "dive into."

My notes:
- Sour means under-extracted.
- My recipe: 18 g in, 36 g out, about 28 seconds.
- When a shot tastes sour, it runs about 18 seconds.
- I chased temperature for a month and tasted nothing.

The reply, in part:

Sour is what unfinished coffee tastes like. Coffee gives up different things at different speeds. The bright, acidic stuff comes out first [CHECK]. The sweet stuff comes next. The bitter, dry stuff comes last [CHECK]. Stop the shot early, and the acids are still the loudest thing in the cup.

Here's how I know mine ended early. My recipe is eighteen grams in, thirty-six grams out, about twenty-eight seconds. When a shot tastes sour, it runs eighteen seconds. Sometimes less.

The [CHECK] instruction is the one I would keep if I could keep only one. A model fills gaps with plausible facts, and in your voice they sound like things you know. Marking them turns your fact check from reading everything into reading a short list.

It used my numbers exactly: eighteen grams in, thirty-six out, twenty-eight seconds, eighteen for a sour shot. It added none of its own. Across all six beats it flagged six things. Three were claims: the order in which flavors extract, the line that temperature moves extraction less than grind does, and the one about old grounds sitting inside the grinder. Three were numbers. In beat 4 it wrote that a one step finer shot might run 22 seconds, the next 26 or 27, and an overshoot 40, and none of those were in my notes. They are plausible guesses, and they are not yours. Replace them with times from your own shots before you say them to a camera.

Beats 3 to 6 took a two line message each: "Now beat 3, same rules. My notes are below." The thread already held the voice, the outline, the hook rounds and my earlier corrections. In a separate chat tab, every one of those messages would have started with a paste.

Read each beat aloud before you ask for the next one. When you rewrite a line in your own words, tell the thread what you changed and why. "I changed this line to X because Y. Match that rhythm from here." That one message teaches it your voice faster than any adjective, and it works on every beat after.

Find out why the draft came in at five and a half minutes

When the six beats were done, I pasted the whole script into a word counter. It was 832 words. At 150 words a minute that is 5:33 of talking, in a video that was supposed to run ten minutes. This is the second thing that went wrong.

The wrong move is to ask for 668 more words, because 1,500 words at 150 a minute is ten minutes, and a model will happily give them to you. What you get is padding in your voice. So I asked the other question, as typed:

The draft is 832 words. At my pace of 150 words a minute that is 5:33 spoken. I need ten minutes. Don't lengthen any sentence. Which beats are thin, and what would I have to bring to fill them?

The reply, in part:

The thin spots are material, not words. Beat 3 says you chased temperature for a month, and the viewer sees none of it. Bring the log: three to five entries with dates, temperatures and what you tasted, on screen for about 45 seconds.

Beats 4 and 5 are demonstrations written as explanations. Film them. Pulling the sour shot, changing the setting and pulling the next two is about three minutes of screen time with very few words over it. Beat 6 promises a channeling video and gives it one sentence. Show a channeled puck or cut the promise.

Rough math: 5:33 spoken, plus 3:00 of shot pulling, plus 0:45 for the log, plus a 0:40 side by side tasting at the end: 9:58.

It was right, and it told me something I did not want to hear. The missing four and a half minutes were footage. This video is part demonstration, and a script that is all words is a script for a podcast.

Put the time budget in the script itself. Write the demonstration beats as bracketed stage directions with a duration, like [PULL SHOT, 0:35, voiceover only], so the runtime you are scripting is the runtime you will film. Your editor will thank you, and so will you on shoot day.

Hand the draft to a different model for a cold read

A model reading its own draft has every reason to like it. It wrote the thing, so it will tell you the second beat could use a little tightening and call that a review. You want a stranger, and the nearest thing to a stranger is a different model family.

In Magai you ask for a different model by name in the middle of the conversation, with a slash pill at the front of the message: /Claude, /GPT or /Gemini. The next reply carries everything said so far, which means the new reader sees the outline, the hook rounds, my notes and every correction without me pasting any of it. Switching models in one workflow covers the mechanics. The short version is that the reader arrives already knowing what the video promised.

In separate tools, the second reader would know nothing. It would not know hook B was rejected for an overclaim, so it might admire the overclaim. It would not know the log is going on screen, so it would flag the missing proof as if I hadn't planned it.

The cold read prompt, as typed:

Read the whole draft as a stranger who clicked "Why Your Espresso Tastes Sour (Fix This First)" and owns a $500 machine. Quote the exact line where you'd reach for the skip bar, the exact line where you'd stop trusting me, and the one line you'd screenshot. One sentence each on why.

I ran it first in the model that wrote the draft. The reply, in part:

Skip bar: "There are five suspects: grind, dose, temperature, puck prep, and the beans. One of them is the cheapest to test and the most likely to be your problem. That's where we start." You told me the answer in the first ten seconds. Now you're pretending it's a mystery, and I know it isn't.

Stop trusting: "Temperature matters, but it moves extraction a little. Grind size moves it a lot." I can't check that, and you haven't shown me anything.

Screenshot: "The good shot lives between sour and bitter, and your job is to find the edge of each and sit in the middle."

All three hits were fair. The tease was a fake mystery, because the hook had already named the grinder. I replaced it with "The first suspect is the one I named in the first ten seconds." The trust problem is the month of temperature logs, and those are now planned for the screen. The screenshot line is the one to say slowly, on camera.

Now run the same message again with /GPT or /Gemini at the front. Compare the two. When two different readers flag the same line, fix it without arguing. When they disagree, you have found a matter of taste, which is yours to decide. I am not quoting a second model's reply here, because yours will flag different lines, and your lines are the ones that matter.

Cut what you'd stumble on, then strip what the chatbot left behind

The last pass has two halves. The first is cutting, and the second is looking for fingerprints.

For the cut, ask for a list and nothing else, as typed:

List every sentence I'd stumble on reading aloud, or that repeats a point I've already made. Quote each one and give its word count. Don't rewrite anything.

It came back with five lines totaling 50 words. I took four of them:

• Twelve words about the water racing through the puck, which the next paragraph says again.
• Five words of "a new set of excuses," which stretched two reasons into a list of three for the sound of it.
• Fifteen words of "you're not there yet, but the water is slowing down."
• Ten words of "one dial is the answer," which repeated the hook with the word the hold test had already flagged.

I kept the fifth, "So here's the video I wish I'd had," because it is the one line in the hook where the viewer meets a person.

That is 42 words, about 17 seconds. The script is now 9:41. It is a ten minute video.

The fingerprint scan is the Hank Green lesson, applied. Scan the finished script for anything that sounds like a reply: [CHECK], "Certainly," "Sure," "Here's a," "I appreciate," "As an AI," "Let me know," and any square bracket you did not write on purpose. Then read it aloud, once, on the pace you will use on camera. Every line you stumble on, rewrite in the words you used when you stumbled. The version you improvised is nearly always the version your viewers will believe.

Last, check every number the model supplied. In this run those were the three shot times in beat 4. A figure from a model is a lead to verify, never a fact to read out.

What YouTube asks of you, and what it leaves alone

Two YouTube policies decide whether you can use AI on a script and keep earning. Both are short, and both have moved before, so read the pages and not my summary.

On disclosure, YouTube's GenAI disclosure page lists "using generative AI tools to create or improve a video outline, script, thumbnail, title, or infographic" among the production assistance creators don't need to disclose. What it does require is disclosure of realistic content that is AI generated or meaningfully altered, such as making it appear a real person said something they didn't. A script you wrote with help is on the leave alone list. Making a guest appear to say something they never said is not.

On monetization, the channel monetization policies say generic or repetitive content includes content that "looks like it's made with a template," and that if you use automated tools or templates to help create your content, the final product must still demonstrate your creative vision. Creator Insider published a walkthrough of that policy in July 2026 with YouTube's VP of Trust and Safety.

Read that with the five asks in mind. Nothing in the process above produces a template. It produces a script built on your notes, your numbers, your footage plan and your corrections, and that is the opposite of interchangeable. The risk is the bare prompt, pasted forty weeks in a row.

Set it up once, then start next Tuesday at the promise

The part you set up once is an agent: reusable instructions that run on any model. Put your voice sample, your banned phrases, your pace, the five asks in order and the [CHECK] rule into it. Building one takes no code, and every thread you open with it starts with your voice already in the room.

You will object that all of this is slower than one prompt, and the first time through, it is. The bare prompt takes two minutes and hands you a script you then have to fix line by line. The five asks take longer once, and after that they are a habit with an agent behind it.

Next Tuesday, you type one line: the idea, plus "start at the promise." The promise, the outline, the hook rounds and the draft run from there, in the order above. The cold read goes to a different family, and the cleanup pass is the same one you ran this week.

When the long video is done, the same thread knows where the best lines are, so cutting the Shorts is one more message and not a second project. The wider picture of where AI fits on a channel is in AI for YouTube creators. If making videos is your job, the Magai page for creators collects what Magai does for YouTubers, newsletter writers and creators. The workspace itself is at app.magai.co.

Ask for the hook. Refuse the first one. Then change one thing and taste it again.