12:56Simon Scrapes · Tutorial
I Rebuilt Hermes in Claude Code (It's Ridiculously Good)
A 13-minute teardown of why rebuilding an agentic OS from scratch beats installing someone else's assumptions.
May 23rdThe argument in one line.
Generic AI output is a context problem, not a prompt problem — three markdown files built once give Claude enough brand DNA to produce personalized content indefinitely without further instruction.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude Code daily and your outputs feel competent but interchangeable with anyone else running the same prompts.
- You create written or visual content regularly and want consistent aesthetics without art-directing each asset manually.
- You have an existing voice from videos, podcasts, or long posts but have never extracted it into a file Claude can reference.
- You are willing to invest 30 to 60 minutes up front in exchange for compounding returns on every future content task.
SKIP IF…
- You are brand new to Claude Code and have not shipped a single agentic workflow yet — this system amplifies an existing workflow, it does not replace learning the basics.
- You do not produce regular content and have no distinct brand identity to encode.
The full version, fast.
Generic AI content is a solved problem once you stop blaming your prompts and start blaming missing context. This video demonstrates a three-file brand system: a voice profile built from structured personality questions and real content samples, a body-of-work document that codifies the seven to twelve ideas you are genuinely known for, and a tokens.json visual identity file that any Claude skill can inject into carousels, slides, and thumbnails. Build all three once and every future output inherits them automatically. As a byproduct, content trained on your actual voice scores around 12 percent on AI detectors versus near-certain detection for generic output.
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 – 00:45
01 · Why brand context matters
Side-by-side cold open: generic AI carousel versus two branded assets. States the three-file promise.
00:45 – 06:50
02 · Pillar 1 — Brand Voice
Live demo of the marketing-brand-voice skill. Four build modes, structured personality Q&A, example collection, voiceprofile.md output. AI detector validation at 12.2%.
06:50 – 09:35
03 · Pillar 2 — Body of Work
Prompt-driven process to extract recurring ideas from past content. Outputs a bodyofwork.md with a core thesis and 7-12 concept pillars.
09:35 – 16:25
04 · Pillar 3 — Visual Identity
Live demo of the marketing-visual-identity skill. Outputs tokens.json, brand guidelines PDF, and LinkedIn carousel templates. Visual references are the critical input.
Lines worth screenshotting.
- Generic AI content fails not because of bad prompts but because Claude has no stored knowledge of who you are — the fix is a file, not a technique.
- 80 percent of brand voice is personality: the stances you take, the things that annoy you about your industry, and the way you naturally explain things to a friend.
- Transcribing your own videos and podcasts produces better voice-extraction data than polished written content, which has already been smoothed of its idiosyncrasies.
- A voice profile is a living file — the way you wrote three months ago is not how you write today, and Claude will drift toward your old self if you stop updating it.
- Splitting your sample guide by platform gives Claude the nuance to write differently for LinkedIn versus a community post while keeping core values constant.
- The body of work is not a topic list — it is seven to twelve recurring ideas, each with a single-line thesis and three strong supporting opinions that underpin it.
- A tokens.json file with colors, fonts, and layout rules lets every downstream Claude skill produce visually consistent output without manual art direction on each piece.
- Content built from your real voice scores dramatically lower on AI detectors — not because it games the algorithm, but because it is actually more human.
- Gathering 20,000 to 30,000 words of your own unguarded writing before running body-of-work extraction is the minimum for reliable pattern detection; below that, patterns are coincidence.
- Visual references are the single most load-bearing input for an identity skill — five strong references produce usable templates; zero references produce a generic brand default.
- Building the three files once compounds: every new skill you add that reads them produces on-brand output automatically, so ROI increases with every piece you create after setup.
One setup session that pays back forever.
WHAT TO LEARN
The gap between generic AI output and output that sounds like you is not about better prompts — it is about giving the model a file to read.
- Claude has no memory of who you are between sessions; a voiceprofile.md changes that by giving it permanent access to your personality, stance, and language patterns.
- Transcribed speech from your own videos or podcasts is the best source material for voice extraction — polished writing has already been edited into genericness.
- Your brand voice is a living document that needs regular updates; the version you built six months ago will slowly pull your outputs toward who you used to be.
- The body of work is not a topic list — it is the seven to twelve ideas you return to repeatedly, each with a core thesis, which means your AI content is anchored to a genuine point of view rather than invented angles.
- Design tokens (a single tokens.json) let every visual task inherit your palette and fonts automatically, removing the need to art-direct each asset.
- Visual references are the highest-leverage input when building an identity system; five strong reference images produce usable templates while zero references produce a default that looks like everyone else.
- Building these three files once compounds: every new skill or workflow you add that reads them becomes brand-consistent from day one with no extra work.
Terms worth knowing.
- Brand context
- The set of files (voice, ideas, visuals) that Claude reads before generating output so it understands who you are without being told each session.
- Body of work
- A markdown document capturing the seven to twelve ideas a creator returns to repeatedly, each with a core thesis and supporting opinions, used to anchor AI-generated content to a genuine point of view.
- Design tokens
- A tokens.json file encoding color palette, typography, and layout rules that Claude Code skills can inject into visual outputs to enforce brand consistency.
- Voice profile
- A markdown file capturing how someone writes and speaks — word choice, sentence rhythm, stance, filler patterns — built from personality Q&A and real content samples.
- AskUserQuestion
- A Claude Code feature that pauses skill execution to prompt the user for a multiple-choice answer or text input before continuing, used here to drive the brand voice interview interactively.
Things they pointed at.
Lines you could clip.
00:00
“The single most important thing that you can do to get better AI content is not better prompts.”
06:10
“80% of your brand voice is dictated by your personality.”
08:50
“The way you wrote three months ago isn't the way that you probably write today.”
14:08
“All your thumbnails, your carousels, your slides, your ad creatives will all be recognizably the same brand without you having to sit there and actually art direct each one.”
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.
story
00:00The single most important thing that you can do to get better AI content is not better prompts. It's adding your brand context. It now has a voice.
00:07It's got a point of view, even a set of color palettes that it would default to. But the problem still is none of it is yours. So on the left, you've got competent AI content that could belong to anyone.
00:17And on the right, two of my own assets that I've created. And you can tell they're mine before you read a word. And importantly, I feel like they're me too.
00:25So the issue is that Claude doesn't know who you are by default. So by the end of this video, you'll have three files to fix that. A brand document like this one, a voice profile so good that you'll wonder why you hadn't done it sooner.
00:36And from that point on, everything you produce sounds and looks like you. So you do the work once, and it pays back for a long time. So let's get started with the voice segment.
00:46So AI content is now sounding quite good. It's competent, but not personal. And even if you don't know it yet, you speak in a very personalized way from the word choices you use, the pacing, the filler words.
00:56And I only came to realize this about myself when I started filming and then transcribing my own videos for my own brand voice. You might have something basic already set up that gives you a few voice rules, but what we need to do is actually extract your values and the way that you think to make it even more personalized and the opinions sound like yours.
01:11Now the brand voice has to live as a file that Claude can read every single time. So we've got voiceprofile.md, which sits inside our brand context folder.
01:19So what we've done is basically turn this process of extracting your brand voice into a skill that you can use today, and it is the marketing dash brand voice skill. And the original came from a member of the Ageneca Academy. So he posted a brand voice playbook that he'd been using for his own writing, and the responses he was getting were ridiculously good.
01:36So I tried it for myself, iterated on it, and then turned that playbook into a skill so that anyone in the community could run it. And I'm gonna run it now as if I'm setting up a new brand voice from scratch and show you a sample of the kind of questions it asks as you go through. And I'll show you where to get these skills for free later on in the video.
01:51So we're gonna run the slash command just to make sure it works, marketing brand voice, and we're gonna say, help me create my brand voice. Now we've got several different modes. We can import existing brand voices.
02:00We can extract it from our raw content, so paste in samples, or we can build it from scratch. So say you're starting fresh or your existing content is already generic and AI generated, and actually what we wanna do is reliably extract from questions that will ask you to pull out your personality. And the last method is actually going and scraping.
02:16So you could just give it access to your LinkedIn URL, a website, etcetera. But you need to make sure that you've got the relevant scrapers already set up.
02:22There. We're gonna start it from scratch and use the ten to fifty minute playbook version, and we wanna go into this full deep dive. So this next concept is really important.
02:2980% of your brand voice is dictated by your personality. So it's the stance you take on things. So it's gonna ask you things like, do you have an anti corporate stance?
02:37Like, what annoys you most about how business is done in your industry? So this is about taking a unique angle and what do you actually represent when you want to talk to your customers or your audience. So I prefer in detail, real examples versus fluffy demos.
02:50When someone comes to you with a problem, how do you naturally respond? So I'm pretty collaborative. It's gonna go and use Claude's ask user questions feature to actually prompt you to answer multiple choice questions or paste in certain bits of samples as you go through.
03:01In real conversations, are you more I'd say I'm gentle but truth telling, but also explain through examples. So I'm gonna say two and three.
03:08And then it starts to ask you things about your actual story. You know, what's your professional background? And you can just extract and pull this from various different places.
03:15But you start to get a picture of how it's building up your personality first. Then if you don't have an ICP or a positioning dot m d file already existing, then it's basically gonna go through and try and understand your strategic framework.
03:27So it'll go through and ask a series of questions about your clients and what you represent or what your brand represents when you're actually marketing to others. So ideal client pain, who are you talking to, and what's the exact pain they're feeling right now that makes them open your email or click your post? And it's really important that throughout this, you obviously bring in as many examples as you can that represent actually what you've done or what you represent.
03:47So minus business owners or aspiring business owners who want to leverage AI and claw code to increase their output, retain high quality outputs, and ultimately save their time doing so. What tone resonates with your audience? What kind of voice do they already trust?
03:59What would immediately tell them off? So SMBs trust non jargon speak, ironic that I said SMBs, and no fluff. When creating content, you feel most authentic when you are sharing personal stories and lessons, breaking down complex topics simply.
04:12Yeah. Let's go with that one, and it gives us a few options that we can go for there. And then natural language patterns.
04:16So how do you actually talk? Do you use we and us or you and I? Do you tell stories or stick to facts?
04:21Any words or phrases so often that friends tease you about them? So you can see how it's really bringing out actual vocabulary phrases, word choice.
04:28And then step three, once we've done the personality and the strategic framework stuff, is example collection. So this is possibly the most important part, which is actually going to your LinkedIn if you feel like that represents you, going to your landing page if you've got content that represents you, if you've got a blog post or used to have a blog post, or just transcribing something directly into here exactly how you would speak.
04:47The best kind of use cases are where you've been on a podcast, you've got a video that you can play back and transcribe of you telling somebody something, then that is the best examples that you can use to actually bring out your accurate voice profile that's gonna sound like you. Or, of course, you can skip this, it'll build from your answers alone.
05:00You'll just get a less good output from the back of it. It will then generate three example voice test samples for an email opening, a LinkedIn post, a landing page headline, and supporting copy.
05:10You can have a read through those, iterate based on that. I wouldn't worry too much about these sounding exactly like you because this is a live file that you'll update over time. And when you apply it to real business context and your examples, that's when you'll start to actually iterate on this and realize how close it sounds to your own personality and the way you speak.
05:26And I'm gonna come back to my agent view here and say, write me a LinkedIn post about ClawCode, and it will go through and actually then generate one. And what I'm gonna do is paste it directly into an AI detector to show you that if something sounds like you, it's less likely to get flagged by something like AI detectors.
05:40Now I'm definitely not encouraging just generating and churning out content. This is about vetting it and retaining high quality, but I just wanna prove that actually something might sound quite like you and therefore not be detected as AI slop or AI content like this. You can also add things like a humanizer skill to make sure that actually we can run it through these and make it even less robotic sounding, but still retain the patterns that you use in your brand voice.
06:03So there we go. We've taken that, and it's basically identified parts of the hook that might be AI or GPT generated, but actually 12.2% chance of it being AI generated.
06:12So it shows you the power of personalizing it your own voice. And, actually, if you've completed this now and you're thinking this does sound like me, and then it's been validated again by not sounding like AI too, that's a really good sign. But it's obviously not necessary.
06:25You might sound more like AI than others, but it's nice validation. So I've actually had mine running for around a year now, and it's the file I update the most. The way you wrote three months ago isn't the way that you probably write today.
06:36And to make this even better, I would highly recommend splitting out your samples guide into individual samples per platform. So you can see I've got things like LinkedIn, school, etcetera, split out here. And that's because on each platform, I speak in a slightly nuanced way, but still retain some of the values and core concepts there.
06:52So that's the end of the first one, voice. And voice is obviously how you sound, and the body of work is what you think and why you think it. So this is gonna eliminate a lot of the work that you have to do when you're talking to AI day to day because it's gonna inherently understand the things that you talk about and the way or the opinion you have on those things.
07:09Things. But before we get to that, YouTube tells me that 97% of you watching this video right now are not subscribed. So head down below and hit the subscribe button if you're getting value from this.
07:18You can also grab the free skills that I've included below in the link in the description. So the body of work, you can think of as the foundational concepts that you find yourself talking about again and again on social media to your friends. So if, for example, you were on a podcast, these would form the foundations of how you'd actually answer the questions you were asked.
07:34And all these thoughts in here should be thoughts that you're continuously developing over time about your topic area or your area of expertise. So this is much easier to write than doing the brand voice document, and I've also given the prompts down below in the links. But we've basically got a find your body of work process here.
07:50It's not a specific skill, but all you need to do is actually prompt Claude to read this document and develop collaboratively with you your body of work. But to skim over this, we're basically going to gather your past content from chats you've had. So from your long form posts, from your video or podcast transcripts, from your sales pages, about pages, anywhere you've articulated your view of certain things.
08:09Once you've done that, it's gonna basically pull out from that content recurring ideas. A good way to also do this is to get Claude to look over your conversation history or GPT to look over your conversation history in the last thirty to ninety days and pull out the recurring themes here. And it's basically gonna give you a single line thesis and then three strong opinions that basically underpin that thesis.
08:29What we're gonna then do is basically optimize those ideas collaboratively with you, and finally, you will get an output with a single markdown file that is basically your body of work that will feed into all of your existing content. And as an example, you can see what mine looks like. Right?
08:42I've got a core thesis, which is core code is the execution layer for all your knowledge work. And then we've got things around skill systems, individual skills. So this is something that if you've watched my videos before, you'll see we talk about continuously.
08:53We've got skills of the IP. So, again, a little bit about skills and what I think about those, And then talking about, actually, underneath everything, it's all just files. It's about architecture.
09:01So these are the things that I talk about again and again on podcasts, in my videos, in my social posts, and they all are my foundational concepts, seven to 12 concepts that form all of my theories, and I adapt these and build on them over time. So the impact the body of work has is it feeds into every bit of content.
09:17So instead of trying to reword angles all the time that Claude or GPT have produced for you on a certain topic and explain them again and again, it's gonna get you to a 60 to 70% version. So when you combine this with the voice profile, it becomes really powerful. So pieces that you create are gonna be anchored to your point of view, and that's why it's so incredibly useful.
09:36So up until this point, we've been talking about everything that impacts stuff that we write about. We decided to take it one step further and wanted a coherent look and feel about our brand. So this is what most people associate with brand.
09:47Right? Every visual asset should feel variable, but still consistent. It should still engage the audience by having that variability, but actually it like our brand feel.
09:55People recognize our post because there's a consistency about them, the colors, the fonts, the way we use graphics, etcetera. And for this, we wanted to reverse engineer best practice. So I actually went out and paid for a branding agency to do this piece of work for me and then reverse engineered that process into a skill.
10:10Now this skill is one available as a paid extra inside our Agenic Academy community. So if you're impressed by the results and you want to get one of these that produce your brand book and feed that then into all of your visual assets, say your LinkedIn posts, your Instagram reels, your short form to long form content, the slides you generate.
10:26If you want all of that, then you can grab that as a paid extra. But if you just wanna actually work out the logic and copy exactly what I've done, I'm gonna show you the logic behind every step. So it's useful to start by showing you what you'll end up with at the end of this.
10:38So you might have things like your color palette. So you can see that we've got all four colors here, including our second accent color that brings things to life. How those colors actually work together in practice.
10:47We've got the typography, so what fonts we use, how they scale, and then more importantly, how those look in some mock use cases. So we've got some dashboards mocked up here, which show the colors on social posts, etcetera, and how they actually might look when they come out the back of it. We've then done things around logos, logo design, the values of the brand, etcetera, etcetera.
11:06So these form your brand guidelines, but it doesn't only happen there. Right? We can't just have that visual PDF brand book for ClawCode to use it.
11:13It outputs all of the fonts as actual files as well as a tokens dot JSON, which basically has everything that we've just defined, including all the color palettes, etcetera, so that Cloudco can use it to lock in designs specifically based on these design tokens. I we're not gonna get variability every single time. It's gonna be consistent in the way feels and looks.
11:32Now at the back of it, you can see how this looks when it actually generates everything in a consistent brand look and feel across different things. So this, for example, is a carousel generator for LinkedIn and Instagram that you can see has that consistent look and feel throughout individual slides, but also the variability.
11:48So when you apply this brand visual design system to specific other skills like a LinkedIn carousel creator, it injects those tokens and makes sure we build templates around your visual identity. And based on the visual references we give it, it will spin up various templates. So let's go through and run it.
12:04So it's marketing visual identity skill. We just tell it to create an identity for us, and you can speed this up if you already have design tokens, for example, in place. It's obviously found my existing one already.
12:15So given a brand name of Simon Scrapes, and it's setting up the mock directory structure, which basically has a folder for templates. So like I mentioned, if we're connecting this to things like LinkedIn carousels, then we're actually gonna create templates from our visual references. So it's gonna ask us for a bunch of visual references that represent our brand and therefore pull it in.
12:31It's also created the visual identity folder where it will store our fonts, our headshots, logos that we often use, and tokens. And you can see right now that tokens is very empty because we've not mocked up any of this. So So all we're doing is effectively getting it to create these files and folders and then feeding them into relevant skills that create specific things like slides.
12:48So we're asked if we want micro labels at the top like date, brand, and role, and it will start to lock these into tokens.json, and we can enter any answers and basically get this customized to your brand completely. And those will all be consistent then because they're pulled from one file.
13:00And what we're doing is just basically defining a configuration file of preferences and styles that it can pull from. Much like if you were doing this work manually, you pull from your assets. That's exactly what it's doing.
13:10Then we move on to reference materials. So we upload three to five visual style references you admire. These can be linked in carousels from people you like, Instagram posts from brands with a strong identity, a website URL with the right vibe, or brand book PDFs you admire.
13:22Like everything, the more you give, the better I can extract palette typography and design moves. This is one of those that you don't skip because if you skip it, your templates are gonna not gonna look anything like you want them to. So for example, when we're creating a LinkedIn carousel demonstration, we put in these artistic style setups, reference one, reference two, three for the slide bodies, four, and five.
13:44So this was all obviously based around LinkedIn carousels. We can split out different designs and different styles there. Off the back of it, it was then able to actually create from the reference something that fairly similar representation.
13:55And this won't be exactly how it produces it every time, but what it will do is take this slide and adapt it based on that. So that's one that's like our body, but we also have things like the hero image. So we've got the reference image there, and we've obviously turned it into a semi template that we can take and adapt.
14:11Now what it would do is actually replace the key elements inside this visual. So it replaced the Claude. It replaced these guys, and it replaced all of the text, etcetera.
14:20Next, we upload our logo. And what we do is just add all of these to the visual references folder and tell it exactly what they are. Will you personally appear in the post?
14:28Like, you can pull an avatar or cover photo, and it will basically pull from those headshots. You saw earlier where it pulled my headshot and added it into a LinkedIn carousel. And you can even add an existing brand book PDF that you can extract from, landing page screenshots, Figma palettes, custom fonts.
14:42Anything you want to personalize it with, you absolutely can. And because we basically gave it no references as an example, it says, okay. We're gonna create a neutral identity now, and we can swap everything in later when we want to change and update it to be our personal visual identity.
14:56Now at the back of it, it's creating a Simon Scrapes brand guidelines. It's given us a mock logo because we didn't actually give it one. You can see it feels and looks fairly similar to the one we had before.
15:06We didn't give it a logo because, actually, these are all the default settings. So because I didn't give it much information, it's basically gone through. And apart from fonts, which which it asked us when you use Geist, it's basically come through and use the same default color palette, etcetera.
15:19But either way, we've got some brand guidelines out the back of it, and now it will be mocking up things like templates for LinkedIn carousels, etcetera, in the background that can then get injected into any of the skill systems that we use with it.
15:30So all your thumbnails, your carousels, your slides, your ad creatives will all be recognizably the same brand without you having to sit there and actually art direct each one of these because all of the information, all of the design tokens, like the masthead to the headlines, the fonts, etcetera, will all get pushed into every visual output that you produce.
15:48So together, you've got three files. You've got voice, which is how you sound. You've got body of work, which are the thoughts and opinions you have and angles on those topics, and then you've got your visual identity.
15:57So you just need to build all of these once, and then every output that call produces from that point sounds and looks like your brand. So you can create this for your personal brand. You can create this for your business brand.
16:06You can for both. And I guarantee you, this is gonna make you stand out in a world of AI slots. So next video, I'm walking you through actual examples of what this system produces in terms of the LinkedIn carousels.
16:16So you'll see for different topics how consistent the outputs get when that foundation of your design system is in place. Thanks for watching.
16:24See you in the next
The Hook
The bait, then the rug-pull.
Better prompts are not the bottleneck. The gap between generic AI output and output that feels unmistakably yours comes down to a single missing ingredient — and in this tutorial, three files fix it permanently.
Named ideas worth stealing.
00:00list
Three-File Brand System
- Voice Profile (voiceprofile.md)
- Body of Work (bodyofwork.md)
- Visual Identity (tokens.json)
Three files that give Claude persistent brand context: how you sound, what you think, and how things should look.
Steal forAny creator who wants consistent AI output without re-prompting brand context each session
01:17list
Brand Voice Extraction Modes
- Import existing brand voices
- Extract from raw content samples
- Build from scratch via Q&A
- Scrape from LinkedIn/website
Four entry points for the marketing-brand-voice skill, ranked by input availability.
Steal forOnboarding flow design for any AI tool that needs user context
06:50model
Body of Work Extraction Process
- Gather past content (20k-30k words)
- Extract recurring ideas
- Find umbrella thesis
- Optimize collaboratively
- Output single markdown file
Five-phase process for distilling years of content into a structured idea document.
Steal forThought leadership positioning, ghostwriting briefs, AI persona setup
How they asked for the click.
VERBAL ASK
09:35product
“This skill is one available as a paid extra inside our Agentic Academy community.”
Soft mid-video mention after delivering significant free value. Offers the free DIY logic path before the paid pitch, which softens resistance effectively.
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.
