Big Idea

The argument in one line.

Claude Projects can replicate a creator's writing style with near-perfect fidelity once given enough past examples as context, but it cannot originate the lived-experience insight that separates a good script from a great one.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You run a content page and already have a backlog of scripts, posts, or captions you could feed an AI as style reference.
You want a concrete, on-screen walkthrough of setting up a Claude Project (knowledge base + custom instructions) rather than a generic explainer.
You're deciding whether to hand off first-draft scriptwriting to AI and want to know realistically how much editing still falls on you.

SKIP IF…

You're starting from zero content and have no past scripts or writing samples to use as project knowledge — the technique depends on having a corpus.
You want a tool comparison — this is a single-tool (Claude Projects), single-workflow deep dive, not a market survey.

TL;DR

The full version, fast.

A creator with 600K Instagram followers and ~15M monthly Reel views built a Claude Project called ReelBot, uploaded a document containing 400 of his past scripts, and wrote custom instructions describing his hook-body-CTA format and tone. Feeding it a news article produces a full 30-60 second script in one prompt, which he rates around 8.5 out of 10. He estimates 70-90% of scripts now come out of Claude usable as-is, with the rest getting human edits to add insights and opinions the AI can't originate on its own — because Claude wasn't given the years of hands-on context that let a human spot a novel angle or comparison. His conclusion: the technique works because he brought deep domain expertise to the instructions; the AI is only as good as what you feed and tell it.

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 – 00:19

01 · Cold open + proof

A Claude-written Reel is shown live at 1.7M views before any explanation, establishing credibility.

00:19 – 01:19

02 · Channel background + promise

Host establishes his track record (600K followers, 400 scripts written, 15M monthly views) and states the video's premise: can AI write as well as him.

01:19 – 02:32

03 · What Claude Projects is

Disclaimer this isn't sponsored, then an explanation of Claude Projects as a way to give an LLM large amounts of persistent context.

02:32 – 03:42

04 · Confession: not 100% AI

Host admits 70-90% of scripts are AI-generated as-is, the rest gets human edits, and previews why later.

03:42 – 04:06

05 · Setup: Claude.ai account + plan

Shows signing up for Claude.ai and using the $20/month Pro plan.

04:06 – 05:42

06 · Creating the ReelBot project

Creates a new Claude Project named ReelBot and writes the project description (30-60s AI/tech Reel scripts, educational + viral balance).

05:42 – 06:51

07 · Uploading 400 scripts as knowledge

Uploads a document containing 400 past scripts as project knowledge, noting it only used 16% of capacity.

06:51 – 07:40

08 · Writing custom instructions

Defines the hook/body/CTA structure and instructs Claude to mimic his tone and style from the uploaded scripts.

07:40 – 08:27

09 · Feeding it a news article

Copies the Microsoft OmniParser article into the chat as source material for the new script.

08:27 – 09:21

10 · Generating the script

Writes a short prompt (relying on the saved custom instructions) and generates the full script in one shot.

09:21 – 10:34

11 · Grading the output

Reads the generated script aloud, praises the hook and plain-language explanation, rates it ~8.5/10.

10:34 – 11:46

12 · Where AI falls short

Explains what he'd add himself, a comparison to an older GitHub project Claude had no context for, and why.

11:46 – 12:26

13 · Human vs Claude context window

Uses a simple bar chart to argue humans currently hold far more lived contextual memory than Claude, though the gap will shrink as context windows grow.

12:26 – 12:40

14 · Other use cases + closing thesis

Mentions using Projects for marketing campaigns, sales pitches, and data analysis, then lands on: AI output quality is bounded by your own domain expertise.

Atomic Insights

Lines worth screenshotting.

A Claude Project loaded with 400 past scripts used only 16% of its available knowledge capacity, meaning far more context could still be added.
Feeding an AI your own back-catalog as project knowledge lets it copy your voice, not just generic good copy — that is the difference between a usable draft and a generic one.
70-90% of AI-generated scripts were usable as final drafts once the project was properly set up with knowledge and custom instructions.
AI can replicate a proven structure (hook, body, call-to-action) reliably once that structure is explicitly spelled out in custom instructions.
The gap between an AI script and a great human-written one is not grammar or structure — it is the missing comparison or insight only lived experience surfaces.
An LLM's output quality is bounded by the domain expertise of the person instructing it — prompt engineering is just clearly explaining what you already know.

Takeaway

Feed the AI your own back-catalog, not just instructions

WHAT TO LEARN

AI-written scripts only sound like you once you have fed it your own past work as reference; the instructions alone are not enough.

Uploading a corpus of your own past writing as project knowledge is what makes AI output sound like you, not like generic AI copy.
Explicitly naming your script structure (hook, body, CTA) in custom instructions gets that structure followed consistently on every generation.
Even a strong AI draft benefits from a human pass that adds a comparison, opinion, or insight the model had no context for.
The quality ceiling on AI output is set by the depth of domain knowledge behind the instructions you give it, not by the AI itself.
Grading AI drafts against your own bar, not against is this good enough, keeps you honest about how much editing still matters.

Glossary

Terms worth knowing.

Claude Projects: A Claude.ai feature that lets a user create a dedicated workspace with uploaded reference documents (project knowledge) and custom instructions, so every chat within that project has consistent context and behavior.
Custom instructions: A saved system-level prompt inside a Claude Project that tells the model how to behave and respond across every conversation in that project, without needing to repeat the same instructions each time.
Hook: The opening line of a script designed to create a curiosity gap so the viewer keeps watching instead of scrolling away.

Resources

Things they pointed at.

05:44toolClaude.ai ↗

05:45toolClaude Projects ↗

09:05toolMicrosoft OmniParser

15:51toolself-operating computer (GitHub repo)

Quotables

Lines you could clip.

00:11

“This video is at 1,700,000 views, and Claude wrote the script.”

immediate proof-of-claim hook, no setup needed→ TikTok hook↗ Tweet quote

19:02

“An LLM or any kind of an AI model is as good as you instruct it. It follows you, not the other way around.”

tight, quotable thesis line→ newsletter pull-quote↗ Tweet quote

20:16

“Prompt engineering is just a fancy way of saying instruct an AI model in English.”

contrarian, deflating framing of a buzzword→ IG reel cold open↗ Tweet quote

The Script

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

00:00I used Claude AI to write an Instagram Reel script, and I'm about to post that Reel in three, two, one. And it's live.

00:09Let's see how it performs. Yo. Check this out.

00:12This video is at 1,700,000. 1,700,000 views, and Claude wrote the script.

00:20Isn't that insane? We started our Instagram page in April 2023 with zero zero followers. And at the time of shooting this video, we're at more than 600,000 followers in about eighteen months.

00:31And in the course of these eighteen months, we've put up about 500 to 600 videos on our profile, 400 of which I had personally scripted. I used to write scripts every single day in order to create videos that was a perfect blend of value provision as well as virality. And a lot of them actually did go viral.

00:49On Instagram, we pull about 15,000,000 views every month only on reels. But that got me thinking.

00:55What if an AI can actually write as good as me or even better than me? Well, turns out it can. In this video, I'm gonna teach you exactly that.

01:04How I personally use Claude's latest feature Claude projects in order to write my Instagram reel tricks. So hit that subscribe button. My name is Shreedev, and you are watching the AI Lab by 100x engineers.

01:19Firstly, big disclaimer. This is not a sponsored video. Claude is not paying us.

01:23I just absolutely love Claude. I think Claude is one of the best LLMs out there in terms of creative writing, so just wanted to put it out there.

01:31That being said Claude came up with a new feature called Claude projects. Claude projects is kind of similar to GPTs. It's basically your LLM but for very specific tasks.

01:41And the best part, you can provide it a huge amount of context. I can basically provide a context worth a 500 pager book, and the model will go through that data every time it generates an output.

01:53And as a result of that, you get much better output, you get much fine tuned outputs, and you get much more narrowed down outputs, more to your liking. Now I'm gonna demonstrate this feature using the example of how I use this feature in order to automate my script writing process, but honestly, it's a lot more than that.

02:09So make sure you stay until the end. Now before I teach you how to do this, I have a confession to make. I use Claude extensively for scripting, but not 100% of the scripts that go out are generated by Claude.

02:20Probably about 70 or 90% is. My team and I definitely make a few changes here and there. Sometimes we insert our thoughts and opinions into the script so that the video is more genuine and it comes from a human standpoint.

02:33But there is a very strong reason why I do this, which I will tell you in a bit, so hang in there. For now, let's look at the process. To use Claude, go to claude.ai and sign up for a new account.

02:42It's free to create an account. Now I'm using a plan called the professional plan, which costs me about $20 a month, roughly thousand 600 rupees, And this is a pretty decent plan that actually gets me what I want.

02:54Now once I'm here, I'm going to hover on Claude on the top left corner, and I'm gonna click on projects. And once I'm on projects, I will click on create new project, And I will call my project real bot.

03:07Alright? Now it's asking me what I'm trying to achieve, which basically means I should describe my project, the goals, subject, etcetera, etcetera. I am looking to write Instagram real scripts that are thirty to sixty seconds long about AI tools, tutorials, and AI news, which is basically what we do.

03:31These videos are educational in nature and strike the perfect balance between value provision and virality.

03:42Alright. I'm just gonna create the project. Once I'm done with that, this is basically where I prompt it, but that's the basic way of doing it.

03:49The better way or the more advanced way of doing it is to add content or context. Now remember when I told you over the last one and a half years, I've written over 400 scripts? I have all those scripts in a document, and this is basically what that document looks like.

04:04It's basically just a dump of so many scripts that I've written over the past one and a half years. And this is going to be context to Claude.

04:13So I will click on the add content button over here. I'll click on upload from device. And even though there's about three to 400 scripts in this file, it's only taken 16% of the knowledge capacity that Claude projects can actually have, which is pretty cool.

04:28Now since I've added my context, I want to provide custom instructions. So I will click on set customs instructions, and I will basically describe in a much more detailed way on what I'm expecting from Claude.

04:40So it basically says instruct Claude how to behave and respond for all of the chats within ReelBot. And this is basically what my prompt is. Let's read it.

04:49You are an expert Instagram Reels script writer for a page that creates content on generative AI, specifically AI news, AI tools, tutorials, and reviews. I've provided a document with scripts that have performed well on Instagram. The first line of each script is the hook, a line meant to capture the viewer's attention.

05:10The remaining lines are the main body, and finally, the last line is a call to action which goes like, for more AI stuff, follow 100 x, which is usually what I say at the end of every single 100 x video. I want you to follow instructions and adopt the style and tonality of writing similar to the scripts in the project knowledge.

05:28So basically, what I'm asking Claude to do is to copy my style of writing and try to sound like me. Now you can go as detailed as you want in terms of your instructions. I'm gonna save it, and we are all set.

05:40Now let's look for a piece of news in order to create a reloar. Now Microsoft launched a new feature called Omnipasser, which is very similar to Anthropix computer use.

05:50It's basically a feature where an AI agent takes control of your computer and can actually use it like a real human being. It can be for code. It can be in order to perform certain actions, whatever you want it to do.

06:01It's a very exciting capability. So what I'm gonna do is I'm gonna copy this entire article, and I will paste it.

06:09There's a little more left, so I will copy that as well, and I will paste this in. And now we have the entire article on Claude.

06:18I will put these in quotes and I will proceed to write my prompt. Now I don't need to get very complex or wordy over here because I've already given it custom instructions.

06:28I just need to tell it what I want. Write a thirty to sixty second Instagram Reel script in a style of writing similar to the knowledge.

06:44Right? I'm gonna click on send prompt, and let's see what Claude gets us.

06:51Alright. So this is basically what Claude got us. You won't believe what Microsoft just dropped in the AI world.

06:58Okay. That I feel is a pretty decent hook. Right?

07:02A hook is basically meant to capture the viewer's attention. You wanna create a curiosity gap in hooks that makes the viewer want to watch the video for a little longer. So when I say you won't believe what Microsoft just dropped in the AI world, naturally, people who are interested in AI are curious.

07:18So I'd say that's a pretty good hook. You won't believe what Microsoft just dropped in the AI world. A new AI model called Omnipasser that's beating g p t four at understanding computer screens.

07:27This is huge. Nice. Think of it like having an AI assistant that can actually see and understand everything on your screen perfectly, every button, every icon, every menu.

07:38The crazy part, when they tested it against g p t four v, OpenAI's vision model, OpenPASSER won by a landslide. It's so good at understanding interfaces that Microsoft has already started using it to enhance their agents. And get this, it's completely open source.

07:54You can find it on Hugging Face with an MIT license, meaning anyone can use it to build better AI tools. What makes it special is how it breaks down screenshots. It just doesn't see the screen.

08:04It understands what each element does and how they work together. It's like giving AI perfect manual for every interface it sees. This is going to change how we interact with computers forever.

08:15Imagine telling an AI, book me a flight, and it actually knows exactly what buttons to click. For more AI stuff, follow 100x. I'd say that is pretty perfect.

08:24I'd actually give that probably an eight, eight point five on 10. It's pretty spot on. So usually in my reels, I try to break down complex things and explain that in very simple languages, so we're not going into complex language in this particular script at all.

08:38It's actually trying to explain this whole concept in very simple language. You know, it could have said something like, it's an omnivision model agent that actually passes your computer blah blah blah, but it didn't say any of that. It just says that it understands what you see on your computer screen, which is amazing.

08:55It got a pretty decent hook. It got a pretty easy to understand body, and it got the perfect call to action, which is for more AS to follow 100 x. I would give this a pass.

09:04But remember earlier when I said that I almost never use 100% AI generated scripts and it's only about 70 to 90%? Here's why. If I were to write or add something to the script, I would definitely make comparisons of previous such models that has been released.

09:19There was a very popular GitHub repo called self operating computer which was a very primitive version of what these particular features actually do today. I would have made comparisons and highlighted how we've started out with a very primitive version of this feature and how far we've come in order to have a pretty decent version of this feature today.

09:36Now why couldn't Claude actually do that for me? Simply because I just did not give it that context. Claude has no idea that there was a model on GitHub that actually came out which actually did that.

09:48So although it's great at generating these pretty realistic sounding scripts, it's not great at simulating insights like a human brain would actually get insights. I have been following the AI space for years. I have years of contextual memory.

10:01I have so many experiences with AI. I have read so many papers. I have tried out so many tools.

10:06I have really played around in this space quite a lot hands on. And as a result of that, I am able to come up with such insights from memory and experience, whereas Claude is simply unable to. But this is a very temporary limitation.

10:21The moment these context windows get higher and higher, Claude will be able to reference and process much more information than it actually does today. And I don't think this day is very far. A couple of years worth of learning, knowledge, and context is a lot of data.

10:37And slowly but surely, we're actually getting to that point step by step. Now this is definitely a great way in order to use Clot, in order to reference a style of writing and replicate that style of writing.

10:50But there are a lot more other ways in order to use Clot projects as well. You can use it in order to ideate and generate entire marketing campaigns. You can use it to generate sales pitches.

11:01You can use it for data analysis. And all you need to do is give it the right instructions, which brings me to my final point.

11:08Claude would not have been able to output a script like this unless I provided it three to 400 of my scripts. It was necessary for me to go through that grind of creating these three to 400 scripts. Now I know a lot of you in the comments would go like, hey.

11:22Can you please provide those transcripts so that we can also try it? And I get why you would want that. But honestly, it won't be of any use to you.

11:29The kind of scripts that I write is very specific to the kind of page that I run. It's not a generic template that you can use for any kind of scripts. It's only for my kind scripts, which ultimately underlines the importance of domain knowledge or domain expertise.

11:45An LLM or any kind of an AI model is as good as you instruct it. It follows you, not the other way around. And what's key in learning how to prompt any kind of a model is to be a domain expert or have domain knowledge in a particular field.

12:00My domain knowledge comes in applied AI and writing, and I was able to apply a combination of that in order to automate a daily process in my life. What is that process that you can automate in your life? What is the domain knowledge or expertise that you have?

12:13That is what you need to find out. Because ultimately, prompt engineering is just a fancy way of saying instruct an AI model in English.

12:22That's all it is. Unless you know what to instruct, you're not gonna get good outputs.

12:26So fundamentally, it still remains important to be good at your craft. I hope this video was useful.

12:32If you liked it, hit the subscribe button, and I'll see you in the next one.

The Hook

The bait, then the rug-pull.

A Reel scripted entirely by Claude is already live at 1.7 million views before the host explains a word of the method, then he opens Claude.ai on screen and rebuilds the exact project that wrote it.

Frameworks