Big Idea

The argument in one line.

Claude only gets better at writing your scripts if you show it exactly where it missed — and a three-prompt feedback loop run after every script automates that correction permanently.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You already use Claude or another AI for YouTube scriptwriting and find yourself making the same corrections every session.
You have a style guide and audience avatar in your Claude project but the output still doesn't sound like you.
You publish multiple videos per month and want the AI's quality to compound rather than stay flat.
You've ever edited an AI-generated script in Google Docs or Notion without feeding those edits back into the chat.

SKIP IF…

You don't use AI for scriptwriting and have no plans to start.
You publish one video every few months — the compounding benefit builds too slowly to notice.

TL;DR

The full version, fast.

Most creators make their AI script edits in Google Docs or Notion — outside the chat — so Claude never sees the gap between what it produced and what the creator actually filmed. The fix is a three-prompt workflow: a collaboration report prompt that compares draft vs. final and classifies differences as Tier 1 (style/voice) or Tier 2 (structure) learnings; a style guide update prompt that absorbs Tier 1 into the project style guide; and a training doc update prompt that absorbs Tier 2 into the structural training document. Repeated every script, Claude drifts progressively closer to the creator's natural voice.

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 →

Chapters

Where the time goes.

00:00 – 00:24

01 · Hook + promise

Pattern interrupt opener, 3-month usage claim, sets up the one missing habit that compounds across every future script.

00:24 – 01:05

02 · The problem: static documents

Audience avatar and style guide alone are not enough. Making edits in Google Docs or Notion keeps Claude blind to its own mistakes.

01:05 – 01:33

03 · The insight: force the diff

If Claude never sees the difference between what it generated and what you filmed, it cannot improve. The solution is forcing it to compare.

01:33 – 02:14

04 · Collaboration report prompt

Paste the prompt alongside your final edited script. Claude surfaces Tier 1 (style) and Tier 2 (structure) learnings in a structured report.

02:14 – 02:53

05 · Absorbing the learnings

Two follow-up prompts: one updates the style guide with Tier 1 learnings, one updates the training doc with Tier 2. Both get re-uploaded to the Claude project.

02:53 – 03:58

06 · Compounding loop + CTA

Each script adds context. The caveat: the system amplifies the human's own writing quality, good or bad. CTA to next video on curiosity mistakes.

Atomic Insights

Lines worth screenshotting.

Style guides and audience avatars tell Claude what to aim for but never tell it where it missed — that gap is why the output stays flat.
Every edit you make in Google Docs or Notion instead of the Claude chat is a lesson Claude never receives.
The collaboration report prompt turns a one-time AI draft into a living feedback signal: Tier 1 for voice, Tier 2 for structure.
A self-improving AI writing partner is not a feature you pay for — it is a workflow habit you run after every script.
Tier 1 learnings update the style guide. Tier 2 learnings update the training doc. They serve different documents on purpose.
The system only compounds on good inputs: if your own writing principles are weak, you are training the AI on weak principles.
Three prompts and two document re-uploads is the entire maintenance cost of a script AI that gets better every session.
The creator's honest caveat — improve your own craft or you'll train Claude on bad habits — is the most important sentence in the video.

Takeaway

Feed the gap back in after every script.

WHAT TO LEARN

AI scriptwriting compounds only if the model sees the difference between what it generated and what you actually filmed — every edit made outside the chat is a lesson that never arrives.

A style guide tells Claude what to aim for; it never tells Claude where it missed. The diff between AI draft and edited final is the missing signal.
Tier 1 learnings (voice, sentence rhythm, vocabulary) belong in the style guide. Tier 2 learnings (segment order, structural architecture) belong in the training document. Mixing them into one document dilutes both.
The three-prompt ritual — collaboration report, then style guide update, then training doc update — costs about five minutes per script and compounds indefinitely.
Re-uploading the updated documents to the Claude project is the step that makes the learnings stick. Without re-upload, the next session starts from scratch.
The system amplifies whatever writing principles it is trained on. Improve your own scripts first, or the AI learns and scales your weaknesses.
A next-video CTA with no subscribe pitch or sponsor ask is a valid low-friction engagement signal — the absence of friction is the strategy.

Glossary

Terms worth knowing.

Collaboration report: A structured Claude output produced by comparing its generated script against the creator's edited final version, categorizing every meaningful difference into Tier 1 or Tier 2 learnings.
Tier 1 learning: A style-level difference between AI draft and final script (voice, vocabulary, sentence rhythm, casual markers) that feeds into the style guide document.
Tier 2 learning: A structural difference (segment order, hook architecture, overall length, section presence) that feeds into the training document rather than the style guide.
Training document: A Claude project document that holds structural rules and example flags for how the creator organizes their scripts — distinct from the style guide, which governs voice and language.
Claude Project: Anthropic's persistent workspace feature where uploaded documents (style guide, training doc) are carried into every new chat session without re-pasting.

Resources

Things they pointed at.

01:33productClaude Scriptwriting Process + Prompts (free) ↗

00:00productAI Scriptwriting App (early access) ↗

Quotables

Lines you could clip.

01:20

“If every time an AI generates a script, you go and make your edits somewhere else, how is it gonna know what it should be doing better the next time?”

Rhetorical question that lands the core problem in one sentence — no setup required→ TikTok hook↗ Tweet quote

01:31

“Force Claude to learn from its mistakes.”

Command-form punchline, stands completely alone→ IG reel cold open↗ Tweet quote

03:40

“Because otherwise, you're simply gonna be training the AI on bad script writing principles.”

Honest caveat that reframes the whole system — contrarian moment in a tutorial that otherwise sells the upside→ newsletter pull-quote↗ Tweet quote

The Script

Word for word.

analogystory

00:00Okay. Don't write another YouTube script using AI until you've seen this. I promise you I'm about to save you a whole bunch of time.

00:07Now I've been using Claude to help me write YouTube scripts for the last three months and today I wanna show you the most important thing that you're probably not doing that's gonna make each script faster to write than the last one. This takes two minutes and you will be kicking yourself you didn't implement this sooner.

00:22Now you might think that giving Claude an audience avatar and a style guide document might be enough to make it sound like you and to make it generate scripts in your tone. But the truth is those two documents are not enough on their own. Now I don't know exactly what your process looks like for writing a script with AI, but I imagine it will be something like this.

00:39There might be some back and forth going over the brainstorm. You might be deciding on the topic. You might be fleshing out the segments of your video piece by piece before asking the AI to just completely generate an entire script for you.

00:50Now I'm not sure exactly what your process looks like, but the point I'm getting at here is you reach a point where you have a fully written script that in some way you have used AI to help you generate. Now what most people will do here is take this AI generation, put it into Google Docs or Notion and make the edits that they wanna make to turn it into a script that's ready to film.

01:09And this is fine. Right? There are always gonna be changes that you have to make to something that's AI generated.

01:14But the problem is by doing that, the AI isn't getting a chance to learn from its mistakes. Think about it. If every time an AI generates a script, you go and make your edits somewhere else, how is it gonna know what it should be doing better the next time?

01:26So the simple solution here is to force Claude to learn from its mistakes. And I'm gonna show you how to do that now with one of the simplest yet most powerful prompts you will ever need for writing a YouTube script. And by the way, to make this easier, you can grab my whole Claude script writing process including the prompt I'm about to show you in the description below.

01:44It's completely free and honestly, this has been a game changer for my channel. So this is the prompt that we're gonna be using. I want you to copy and paste this exact prompt into your Claude chat.

01:53Now what it's gonna do is look for differences between the portions of the script or the whole script, if that's what you used it for, that it generated compared to the final version of the script that you have edited and thrown back into the chat. That's the key thing here. Make sure you've pasted back in the final version of your script so that Claude has something to compare its own generations to.

02:14Put it in alongside this prompt, and as you can see, what it's gonna do is look for structural changes, segment level changes, style language patterns, all of this, and convert these into what I call tier one and tier two learnings. All of this with the goal, right, that the next time you sit down to write another script using Claude, it's gonna have this continuously updated context on the way that you write.

02:36So run that prompt and that will give you your collaboration report. Next up, and I still get excited that this is even possible, we're gonna ask Claude to absorb those learnings into its own training. So let me show you how to do that now.

02:48So this is the first of the two prompts that we're gonna use to update our two documents. Uh, this one is to update the style guide. So you'll simply copy and paste this in.

02:57It's going to generate another doc with the updated changes to your style guide, you can then simply re upload back into the project on Claude and it's gonna have an updated style guide for the next script. Then finally, we have the training doc update. Now once again, exact same principle, copy and paste this into Claude it's going to update those key structural changes to the training document.

03:18So once again, on your next script, it's gonna know better how to structure out script in a way that you would. Everything the AI learned about how you write in comparison to how it thought you wrote has been absorbed back into its own training document. So on the next script, it's got more context.

03:33And as you keep repeating this process script after script, it's gonna get closer and closer to your style each time. Now the key to this process was forcing Claude to learn from its mistakes. But it's important that you're doing the same too.

03:46Right? Because otherwise, you're simply gonna be training the AI on bad script writing principles. So watch this video next to understand the huge mistake most YouTubers make when adding curiosity to their scripts and how you

The Hook

The bait, then the rug-pull.

There is one step almost every AI-assisted scriptwriter skips — and it is the step that would have made every script after the first one faster to write. George Blackman's answer is three prompts and two document re-uploads, run once after every video.

Frameworks

Named ideas worth stealing.

01:33list

Three-Prompt Feedback Loop

Collaboration Report Prompt — compare AI draft vs. final script, extract Tier 1 + Tier 2 learnings
Style Guide Update Prompt — absorb Tier 1 (voice/style) learnings into the style guide document
Training Doc Update Prompt — absorb Tier 2 (structural) learnings into the training document

A post-script ritual that feeds the gap between AI output and edited final back into Claude project documents, compounding style accuracy across every future script.

Steal forAny AI-assisted writing workflow — replace script with newsletter, email sequence, or blog post

02:14model

Tier 1 / Tier 2 Learning Classification

Tier 1: style, voice, language patterns, sentence rhythm → updates Style Guide
Tier 2: structural changes, segment order, hook architecture → updates Training Document

Separates subjective voice corrections from objective structural corrections so each type flows to the right document.

Steal forAny prompt engineering or AI workflow where style and structure need separate treatment

CTA Breakdown