21:44Luke Scripted · Tutorial
how i write youtube scripts with claude in under 1 hour
A live 21-minute walkthrough of a nine-step Claude scripting system that replaced five hours of blank-page work with one hour of guided conversation.
April 20thThe argument in one line.
A raw Claude prompt produces logically correct but retention-killing YouTube scripts; the fix is a two-layer system of task-level Skills and a brand-level Project, tuned over time with a taste-log feedback loop.
Read if. Skip if.
READ IF YOU ARE…
- You write YouTube scripts for a business or personal brand and have tried AI tools but found the output too generic to publish.
- You have used ChatGPT for scripting, hit the same recurring tells, and want to know if Claude is actually different.
- You are a coach, consultant, or agency owner using YouTube to acquire clients and need scripts that hold retention and drive action.
- You understand Claude basics but have not built a systematic writing workflow around them.
SKIP IF…
- You want a finished script produced for you; this is a system-building tutorial requiring your voice dumps and editorial taste as inputs.
- You make entertainment content where retention dynamics differ significantly from educational or sales-oriented YouTube.
- You have no interest in iterating prompts over time; the taste-log loop requires deliberate effort across multiple sessions.
The full version, fast.
Claude fails YouTube scripts not because it writes badly, but because it writes logically, and logical structure kills retention by giving away the lesson before the viewer is curious. Out of the box, Claude states the answer upfront, uses broad platform vocabulary instead of niche terms, and loses hook rules when prompts grow complex. The fix is a two-layer architecture: task-level Skills that encode how each section type should be structured, plus a brand-level Project loaded with voice guides, audience docs, and a brand bible. A taste-log feedback loop, pairing bad AI output against your rewritten version and feeding examples back to update the skill, makes the system self-correcting and converges it on your voice over time.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:29
01 · Cold open
High-stakes hook: can Claude write scripts that close high-ticket clients? Credibility established with four years as a professional scriptwriter.
00:29 – 01:52
02 · The ChatGPT problem
Two years of ChatGPT tells, em dashes, 3 to 5 hours of rewriting per script. Decision to give Claude a real two-week test.
01:52 – 03:22
03 · Problem 1: Claude kills retention
Claude states the lesson upfront. Miro diagram contrasts the Claude default path with the YouTube-native hook to story to lesson sequence.
03:22 – 04:41
04 · Fix 1: Claude Skills
Script mapped into four parts with one skill per section. Voice dump then run skill then get draft then tweak then update skill loop.
04:41 – 05:51
05 · Problem 2: Wrong language
Skills tell Claude how to write but not who it is writing for. Niche vocabulary gap: engagement versus watch time.
05:51 – 07:23
06 · Fix 2: Claude Projects
Brand bible, voice guide, audience doc, and CTA offers loaded into a Project. Claude applies the files every session.
07:23 – 08:55
07 · Problem 3: The hook keeps breaking
Section hooks either went missing or leaked the lesson. Root cause: too many instructions in the body skill buried the hook rule.
08:55 – 11:09
08 · Fix 3: Taste log and separated hook prompt
Taste log accumulates bad Claude output paired with user rewrites; examples fed back update the skill. Hook pulled into its own separate prompt run after body is complete.
11:09 – 12:10
09 · Where the system is now
60 to 90 minute scripts. Thinking at the gym, voice dump when seated, brain-dump into Claude, one or two word tweaks, done.
Lines worth screenshotting.
- Claude defaults to logical script structure, stating the lesson upfront, which removes curiosity and kills YouTube retention before a section even lands.
- On YouTube, every section hook should tease the result or pain without revealing the lesson; Claude inverts this by default, giving the answer before the viewer is invested.
- Words like engagement and leverage your platform signal AI authorship to YouTube viewers; niche-specific language like watch time and click-through rate reads as authentic.
- Claude Skills encode how to write each section type; a Claude Project encodes who you are writing for; both layers are required and they do not overlap.
- ChatGPT custom GPTs routinely ignored uploaded files; Claude Projects apply context documents every session, so the same strategy can fail on one platform and work on another.
- When a single AI prompt accumulates too many structural variations, important rules like hook formatting get buried and stop being followed.
- Splitting a complex prompt into two focused ones, body first and hook separately, eliminates the hook-breaking problem without any skill rewriting.
- A taste log pairs the AI output at the top with your preferred rewrite below it; accumulate enough examples and Claude can analyze the pattern and update its own skill instructions.
- The self-correcting feedback loop means the system improves with use; each taste-log cycle moves Claude closer to your voice without manual prompt engineering.
- Scripts that required 3 to 5 hours of ChatGPT rewriting compress to 60 to 90 minutes under a properly tuned two-layer Claude system.
The AI script that sounds human needs two layers
WHAT TO LEARN
Claude writes logically correct scripts by default, but on YouTube, logical structure destroys retention by giving away the answer before the viewer is curious enough to want it.
02The ChatGPT problem
- The recurring AI tells that trained readers recognize, em dashes, the not-X-but-Y pivot, result from ChatGPT optimizing for plausibility rather than human voice.
03Problem 1: Claude kills retention
- Every section of a YouTube script should open by teasing a result or pain, build through a story or analogy, and land the lesson at the end; Claude defaults to the inverse, stating the answer first.
04Fix 1: Claude Skills
- Mapping the script into four distinct section types and building one dedicated skill per type eliminates the need to re-explain structure in every session.
05Problem 2: Wrong language
- Generic platform language such as engagement and leverage your platform signals AI authorship to viewers; niche-specific vocabulary like watch time and click-through rate reads as authentic.
06Fix 2: Claude Projects
- Claude Projects apply uploaded context files every session, where ChatGPT custom GPTs historically ignored them; the same strategy that failed on one platform may now succeed on another.
07Problem 3: The hook keeps breaking
- When a single prompt grows too complex with multiple structural variations, the rule you care most about gets buried; split the prompt into body-first and hook-second to restore compliance.
08Fix 3: Taste log and separated hook prompt
- A taste log, pairing the AI bad output against your rewritten version, gives the model the contrastive examples it needs to update its own instructions and converge on your editorial voice.
- Scripts that required 3 to 5 hours of ChatGPT rewriting can reach 60 to 90 minutes with a properly tuned two-layer system, but the tuning requires deliberate taste logging across sessions.
Terms worth knowing.
- Claude Skill
- A saved set of instructions in Claude that executes a specific task the same way every time it is invoked, eliminating the need to re-explain the task in each session.
- Claude Project
- A persistent Claude workspace where uploaded files such as brand guides, voice docs, and audience profiles are loaded into context automatically at the start of every conversation.
- Taste log
- A document that pairs an AI-generated output the user rejected with their own rewritten version directly below it; accumulated examples are fed back to Claude to update the skill that produced the original.
- Voice dump
- An unstructured spoken or typed brain-dump of raw ideas that serves as the human input to a Claude skill run, preserving natural language patterns before the AI shapes them.
- Section hook
- A single opening sentence for a YouTube body section that teases a result, pain, or story without revealing the lesson, designed to give viewers a reason to keep watching.
- AI tells
- Recurring linguistic patterns that signal AI authorship to trained readers, such as the not-because-of-X-but-Y construction and em dashes used as rhetorical pivots.
Things they pointed at.
Lines you could clip.
00:51
“If your viewers know what AI sounds like, and they probably do, your scripts are cooked before anyone even watches them.”
03:00
“It's kinda like a magician who explains the trick before doing it.”
08:11
“The real problem wasn't that the skill was bad, it was that I had no system for making it better over time.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
analogy
00:00Can Claude write a YouTube script that's actually watchable? But more importantly, can it write the kind of scripts that gets your business high ticket clients? Well, I've been writing YouTube scripts full time for four years at this point.
00:12It's literally my job, And and I've written scripts for big channels and business owners as well. So I decided to go all in on Claude for the last two weeks. And I found three main problems that will make your script sound like everyone else's AI slop, and I'll show you exactly how I fixed all three problems.
00:29But first, here's what pushed me to finally give Claude a proper test. So I'd been using ChatGPT for my own scripts for a couple of years at this point, and honestly, it was never that good. So I'd put in a prompt and I'd something back that was technically okay, but it was full of ChatGPT tells.
00:46It was always the same patterns like this, for example, not because of x, but because of y. But here's the truth, and there'll be em dashes everywhere, and anyone who's used ChatGPT for more than a week knows exactly what I mean.
00:59And if your viewers know what AI sounds like, and they probably do, your scripts are cooked before anyone even watches them. So anytime that I sat down to write a script, it turned into the same process. I generate some cringe, and then I'd rewrite it.
01:13I'd have to fix the hook, fix the parts that sound like a robot trying to sound human, and it would take three to five hours every single time. Now, I actually did try Claude once before. It was about a year ago, but I didn't think too much of it, and so I went back to ChatGPT since I was more familiar with it.
01:29But over the last few months, I kept seeing Claude everywhere online. There was Claude Code, Claude Cowork, people saying it was going to replace ChatGPT. And because I never gave it a real test, I figured, alright, let's just give it two weeks and just go all in properly this time.
01:44I wanted to see if the hype was real, or if it was just another AI tool that was being glazed on Twitter. And the first script that came back for me told me quite a lot. And so the process that I went with was this.
01:54I sat down, and I did a little voice dump, where I just dumped out my ideas, and then I told Claude to write me the intro and the first section of my video. But the first thing that I learned is that Claude has no idea what an actual YouTube script is. Now the problem is that it opens every section by just saying the key point.
02:12So it states the lesson upfront, and there's no curiosity created when they do this, and so the viewers will drop off. Now this is logical, but on YouTube, the structure kills retention. So to show you an example right here, as you can see here in the first line, it says, the most important thing to understand about writing hooks for YouTube is that your hook needs to clearly communicate the main lesson of the section upfront.
02:32So the issue here is it just gives away the lesson. Right? Now what you're gonna do is start each main body section with a hook.
02:37And the whole reason you do that is because the hook creates a reason to keep watching. And so the way I like to do things is to use a hook to tease the result, and then I build into it with a story or an analogy or something like that. And then from there, then the lesson lands at the end when the viewers earned it.
02:53But Claude makes the mistake where it just skips straight to the answer, and that ruins all the curiosity. It's kinda like a magician who explains the trick before doing it.
03:01Right? So, for example, here's a slightly better version. You see how here, we start with, here's the one thing I changed that made my scripts go from getting clicked off in thirty seconds to actually holding viewers through to the end.
03:12Here, we just tease the benefit. There's no leaking of what the main point is, and then we dive into a story right here, and then the lesson that came from it. And so this is more of what we want to get.
03:22And so after messing around with Claude, I found out about something called Claude skills. And so now, instead of having to explain what I want every single time, all I had to do was build out the instructions once, and then save it, and then Claude would run from that every time that I used the same skill. And so what did was I mapped out the process into four parts.
03:40The outline, the intro, the main body sections, and then the CTAs. And then from there, I made one skill per section. And so as you can see right here, we have all the Claude skills that I used.
03:49Got Miro visuals, CTA writer, hook writer, outline writer, all this stuff. And so anytime I needed to write an intro, for example, Claude would just go, take the Claude skill, and then use that to help me write out the intro. And so now I don't have to tell it how to do it every single time I write.
04:04So then the process I decided to start taking was I would just voice dump the idea, run the skill, and then I'll get a draft back, I'll then make some tweaks, and then I'll then adjust the instructions over time so that the skills actually matched what I wanted. And when I started doing this, Claude stopped writing article looking YouTube scripts, and it started writing scripts that were optimized for audience retention.
04:25Now I put all of my skills right here for script writing in the description below this video. It's in a Google Doc, and it has all the skills that I'm currently using, and I'm updating them every time I make changes. And so if you wanna grab these skills for free to write your YouTube scripts in half the time, then check it out.
04:40The link is in the description. So now that the structure was fixed, but reading through the drafts, I still felt like something was wrong, because Claude kept using words that nobody in my niche actually uses. And so I help business owners write YouTube scripts that gets them more sales, but Claude would use these weird words like engagement, or content pillars, leverage your platform.
05:00These are just broad social media terms. It's kind of the thing that you might see on a LinkedIn post from some marketing consultant or someone like that. But if you're talking to a YouTuber, you'd probably say words like watch time, audience retention, click through rate, views.
05:14Like, these are words that are actually specific, and unlike a word like engagement, they actually mean something to my viewers.
05:20And so when your scripts are full of words like engagement, and leverage, and things like that, your viewers would just kinda feel like the script is off. And maybe you can't pinpoint why it feels off, but it just sounds weird, and that's how you get that AI slot feeling in your videos. And so because the skills told Claude how to write, it didn't really tell Claude who he was writing for.
05:40And so things like my voice, my niche, the specific language that my audience actually uses, that was something that Claude wasn't able to do. And so I needed to fix this issue next, and it needed a different kind of solution. So the fix was something I'd actually tried before, but I wrote it off because it didn't work that well.
05:56So Claude has a feature called projects, as you can see. And it's basically a workspace where you can drop in files, and Claude reads them every single session.
06:04And so I put in the brand bible, some voice docs, CTA offers, the target audience, basically everything Claude needed to understand my brand before it writes a single word. Now I tried something similar in chat GPT with the custom GPTs where you can give them files, but didn't quite work that well. Like, it would just ignore the files, didn't really follow them properly, and so I just forgot about it and stopped using it.
06:27And so I was thinking with Claude, it might do the same thing, but then also it's a few years later, so the tech might be better now. And so I decided, why not try it? And so I added these files right here, and it made a real difference, actually.
06:37Like, it wasn't perfect straight away. There was still some issues, but there was way less corrections that I had to make, and Claude started using the right language. And so for most sections, I'd just do a voice dump, as you can see right here.
06:49And then Claude would write it out. And sometimes I just copy and paste the whole thing into my Google Doc without changing a single word. So quick distinction between skills and the Claude project.
06:58So skills are for the task level. So this is things like how to write a hook, how to write the main body section, things like that. Now the project, on the other hand, is at a brand level.
07:07So this would be who am I, who am I talking to, how do I speak, things like that. And so skills work across all my channels, whereas the project is specific to one YouTube channel.
07:18And so you need both of these to be able to hit your voice, as well as how to actually write properly. So at this point, the scripts were coming together pretty fast.
07:25I just do the voice dump, and then Claude writes it out, and then copy and paste, and then we're done. And this happened most of the time, but there was still one thing that kept breaking. Because I couldn't quite get it to the point where I could just copy and paste what Claude gave me every single time.
07:38Still had to make some changes. It might be like a word or two here and there, but the main part it just couldn't get right was the hooks. Because every main body section is supposed to open with a hook.
07:48You just want one sentence that teases the result, or the pain, or the story without giving away the main point. It's how you're able to get people to keep watching. But Claude would either just skip the hook entirely, or it would write a hook that gave away the lesson.
08:00So And this was the same issue I was trying to fix at the start, and it just kept showing up over and over again. So I tried adjusting the skill. I rewrote the main body section skill three or four times, but it still just kept breaking.
08:10I tried adding examples, different rules, and things like that, but it just couldn't seem to get the hooks right. So the reason I think this was, is because I think the prompt was just too big.
08:20Like the skill for main body sections, there were so many variations, like sometimes it's a story, sometimes it's an analogy, sometimes it's just a framework. And so because there was too many instructions, the hook rule probably just got buried inside of the prompt, and so I figured out a solution. The quick thing before we get into the fix, I'm launching a new one on one coaching program where I work with you directly on writing YouTube scripts that get sales for your business.
08:44And building up this AI system is a big part of it. I only have three spots open for April though, and once they're gone, they'll be gone until I open them again. So if you want in, the link is in the description below this video.
08:55Now the fix for the hooks in our scripts came down to two changes. The first one was building a feedback loop because the real problem wasn't that the skill was bad, it was that I had no system for making it better over time, and so I built one. So I started building what I call a taste log.
09:11And so every time Claude produced the section that I didn't like, I'd save it, and then I'd write down my own version right after it. And so Claude's version would be at the top, and then my version on the bottom right below it. And once I had enough examples, I'd then feed the whole thing back into Claude and say, here's what you gave me, and here's what I actually want.
09:30Can you look at the patterns, and then update the skills to match more of what I want? And so here's what I mean by a taste log. I'm actually going run through the process of what I would do to fix the skill.
09:40So we have here four sections that we have Claude's version, and then my version. So I would just say, help me update my main body skill. So then I'd say, here are some examples of Claude's version, and my final version.
09:51And so Claude would just read through the examples you give it, analyze the differences, and then it rewrites the skill instructions for you. And so all you gotta do really is just download the new version, and then replace the old version of the skill. And just like that, we have the new version of the skill created right here.
10:07So we just click save skill. It then says upload and replace. You click that, and then the new skill is created.
10:13And so there we go. Here's the full skill. I'm actually gonna update the skill inside of our Claude skill bank right here.
10:21This again, link is in the description if you want access to this. But I have all the skills that I use to write my scripts. So let's just paste the new version here, and there we go.
10:28It's now updated. And so we have this improvement process in place for the skills, but we still had another issue, because the hook just kept breaking. And so what I did here was I just pulled the hook out entirely, and made it so that, instead of writing it with the main body section, we would do the main body section first on its own, and then we'd do the hook afterwards.
10:48And so, all I do is I just copy and paste the full main body section, and I say, hey, can you take a look at this and give me five hook ideas, please? And then it writes out five hook ideas, I then pick the best one, and just add it in at the start of the section. And so that was it, really.
11:02We have the feedback loop right here, then we separated the hook prompt from the main body prompt. And with this system in place, everything started working pretty good. So here's where I'm at right now.
11:11Now I'm able to get grips done in sixty to ninety minutes per script. Now, I'm a bit of a perfectionist, so I'm probably not the fastest benchmark. But if you move through it without overanalyzing every line, then you could probably get it done in an hour, maybe less once the system is dialed in.
11:25So the workflow is pretty simple now. Most of my thinking happens when I'm away from my desk. So I'll be at the gym, and maybe an idea comes out, and I'll jot down a note in my phone.
11:34And then later in the day, when I sit down, I've already got a rough outline from just thinking it through in the gym. And so from there, I just start brain dumping into Claude. I just go back and forth, we chat basically, it acts as a sparring partner, and then, all I'm doing really is just refining it, changing one or two words, and then we're done.
11:52And so now you've got a system that writes scripts in sixty to ninety minutes without sounding like everyone else's AI slop. But the next problem is knowing what to actually make videos about so the right people find them. So watch this video next on how to gaslight the YouTube algorithm into feeding you more views.
12:07I'll put the video up on the screen for you to watch next.
The Hook
The bait, then the rug-pull.
Four years of writing YouTube scripts professionally, two years of fighting ChatGPT slop, and a two-week all-in experiment with Claude that produced a system, not just a verdict.
Named ideas worth stealing.
03:22model
The Two-Layer AI Writing System
- Claude Skills (task level: how to write each section)
- Claude Project (brand level: who you are, who you write for, your voice)
Skills alone produce on-structure but off-brand output. The Project alone lacks structural rules. Both layers together produce retention-optimized, on-voice scripts.
Steal forAny AI-assisted writing workflow targeting a specific audience or channel
08:35model
The Taste Log
- Save Claude bad output
- Write your preferred version directly below it
- Accumulate examples
- Feed pairs back to Claude
- Ask Claude to analyze the difference and update the skill
A self-correcting feedback loop that converges the skill on your editorial voice without manual prompt engineering.
Steal forAny AI workflow where output quality drifts from personal taste over time
02:22model
Hook-Tease to Story to Lesson
- Hook teases the result or pain without revealing the lesson
- Story or analogy builds into the point
- Lesson lands at the end when the viewer has earned it
The retention-native structure for YouTube body sections; inverse of Claude default behavior.
Steal forYouTube scripting, email sequences, any long-form content requiring section-level retention
How they asked for the click.
VERBAL ASK
08:35product
“I'm launching a new one-on-one coaching program where I work with you directly on writing YouTube scripts that get sales. I only have three spots open for April.”
Embedded mid-video as a before-the-fix aside; landing page shown on screen with 30-day Script-to-Sales framing.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
