§ 01 · The Hook
The bait, then the rug-pull.
Within six seconds Zinho plants a claim so clean it doubles as a thesis: Claude is the strongest AI available in 2026 — not because of one feature, but because of the combination. Every chapter that follows is a brick in that argument, stacked in order until the final demo proves the whole system fires from a single sentence.
What the video promised.
stated at 00:56 "I'll show you exactly how to build your very own [system] in this video." delivered at 19:33
Where the time goes.
00:00 – 01:19
01 · Cold open + three pillars
Cinematic title-card hook. Claude positioned around three pillars: assistant, Code, Cowork. Single-prompt script demo teased.
01:19 – 02:07
02 · Step 1 — Pick the model
Sonnet for 90% of tasks, Opus when the answer matters more than speed, Haiku never. Toggle web search on — most complaints about outdated answers trace to this switch being off.
02:07 – 05:15
03 · Step 2 — The prompt rewrite
Google trained us to ask short questions. Claude needs context like a new hire. 3-part formula: role + context + output shape. Interview flip: end with 'before you write anything, ask me questions.'
05:15 – 06:24
04 · Step 3 — Memory
Claude has account-level global memory. Import trick: use the in-app prompt to export GPT/Gemini history, paste back — years of context ported in ~5 minutes.
06:24 – 09:04
05 · Step 4 — Artifacts
Artifacts build interactive tools directly in the chat panel. Demo: ROI calculator built in 90s, styled via chat feedback, published and shareable. File uploads (PDF/CSV/screenshots) feed into artifacts for analysis.
09:04 – 11:01
06 · Step 5 — Projects
Persistent workspaces with files, instructions, and conversations. Upload brand guidelines + voice sample once — every conversation in the project reads them permanently.
11:01 – 14:31
07 · Step 6 — Skills
Instruction files Claude reads when a task matches. Write once, fires automatically. Demo: LinkedIn post skill described in natural language, Claude generates the file, one-line debug method shown.
14:31 – 16:44
08 · Step 7 — Custom instructions
5-section 200-word briefing: who am I, how I want Claude to talk to me, what I'm working on, how I work, what to never do. Before/after demo: same model, wildly different tone.
16:44 – 19:33
09 · Step 8 — Connectors + Research
Connectors plug Claude into Google Drive, Calendar, Gmail, Notion, Slack — pulls docs on demand. Research mode: autonomous multi-search with citations (5–15 min). Extended thinking for strategy/finance.
19:33 – 20:20
10 · The whole stack fires
Single sentence typed inside the YouTube scripting project: custom instructions + memory + skill + Drive connector all activate. Output is a filmable draft in his voice.
20:20 – 20:59
11 · Outro + CTA
Claude Code and Cowork walkthroughs teased. Free School community CTA.
Visual structure at a glance.
Named ideas worth stealing.
03:01
list
The 3-Part Prompt Formula
- Role (specific)
- Context (situation + what worked)
- Output shape (format/length/tone)
Minimum viable brief for any Claude prompt. Without all three you get Google-quality generic answers.
Steal for JoeFlow/MCN onboarding prompt templates
02:38
concept
The New Hire Reframe
Claude is good when treated like a new hire. Bad when treated like a search engine. Immediately fixes most prompt failures.
Steal for Cold-open for any AI tutorial
04:38
concept
The Interview Flip
Append 'before you write, ask me questions you need' to any prompt where context is unclear. Claude becomes the interviewer.
Steal for Standalone JoeFlow tip; zero setup
15:05
list
Custom Instructions 5-Section
- Who am I
- How I want Claude to talk to me
- What I'm working on (update weekly)
- How I work
- What to never do
Tight 200-word universal briefing. The 'what to never do' section has highest ROI per word.
Steal for Template to ship with JoeFlow or MCN+
12:01
model
Skills vs Projects vs MCP
MCP = kitchen (connection layer). Projects = workspace. Skills = the recipe sitting on top of both.
Steal for Explainer graphic; parallels CLAUDE.md + slash commands
13:28
concept
Skill Debug One-Liner
Type 'when would you use the [X] skill?' — Claude quotes the trigger description. You see exactly what's vague. Edit one line, reinstall.
Steal for Include in any skills tutorial
Lines you could clip.
02:38
"Claude is good when you treat it like a new person that you are hiring for your business. It's bad when you treat it like Google."
10:19
"It's not a better prompt. It's basically a better environment."
06:08
"The single highest leverage hour that you'll spend on your AI setup this entire year."
04:38
"Claude knows what it needs better than you do."
16:06
"This is a briefing. Keep that in mind. It's not a biography, so you don't need to give it so much."
How they spent the runtime.
Hook length56s
Info densityhigh
Filler8%
Things they pointed at.
14:13toolAnthropic skills directory
14:14toolNotion skill
14:15toolFigma skill
How they asked for the click.
20:20
next-video
"The templates, my custom instructions, the exact skills that I use, and the rest of my workflows are all in our free school community."
Soft. Free community link in description. Two follow-up video CTAs teased. No hard ask.
Word for word.
HOOK opening / re-engagementCTA the pitch analogy
00:00HOOKYou've probably seen Claude everywhere lately, and I'm gonna say something that might sound like hype, but I genuinely mean it.
00:07HOOKClaude is currently the strongest AI tool available to most people in 2026, not because of just one feature, but because of the combination. Claude is built around three pillars.
00:20HOOKClaude itself, the assistant which we are going to be using today, Claude code, is for shipping software, and then Claude co work which is the agentic layer that runs tasks on your computer while you are doing something else.
00:37HOOKNow I want you to watch this. I can just type in one sentence to offer me a video about why most people get bad results from AI tools, and then Claude pulls in my voice, my audience, my hook formula, my structure introduces a script that I'd actually film, and all it took was just one prompt.
00:56HOOKThe whole system fires just from using that one prompt. Now that doesn't happen out of the box. There's a full setup, and I'll show you exactly how to build your very own in this video.
01:07HOOKAnd also, I've got dedicated walk throughs on Cowork and also Claude. Those are linked inside our free school community. The link is down in the description below.
01:18HOOKSo we are starting off with step number one that is picking the model, turning on web search, and basically so that you know what you are paying for. Okay. So there are three set of decisions and they take about thirty seconds each.
01:29So the first one is the model. So Sonnet is what I'm gonna be using today. It's fast, it's sharp, it handles about 90% of what you'll throw at it.
01:36And then Opus is what I switch to when I need Claw to actually think. So that's more for strategy and research. So anything where the answer matters more than the speed, if you get what I mean.
01:48And on that note, Haiku, I never ever touch it. And then the second part is web search. Okay.
01:52So in settings, I'm gonna toggle it on and then make sure to leave it on and half the claw doesn't know recent stuff, complaints that people are actually filling out these days. Literally, it's because they never ever flipped on the switch. Now we get on to step number two and that is the prompt rewrite.
02:08So I typed in, write me a content plan for my YouTube channel, and here's what it came back with. Consistent posting schedule, know your audience, mix educational and entertaining.
02:18And, I mean, you could have googled this all the way back in 2015, and that was eleven years ago. Now Claude isn't isn't necessarily bad, it's just that the prompt here is bad.
02:29So Google trained us to ask short questions and to expect good answers. Now that works for search, but when it comes to Claude, it doesn't necessarily work. So what I would say is the way you should reframe this, it will change everything.
02:44So Claude is good when you treat it like like it's a new person that you are hiring for your business. It's bad when you treat it like Google. Now if you hired someone on Monday and on day one, you said write me a content plan, what would they do?
02:56They produce exactly what Claude just produced, uh, to be generic, to be pretty basic. But to change it up, you can brief them on who you serve, what you've tried, and also what is working.
03:08You want to give them something that is usable. Now it's the same thing here. Make sure that you have three things in every prompt.
03:14So number one, the role. So what is Claude operating as? Not you are the expert because that's pretty vague.
03:21You are helping me to build a YouTube channel in the AI tool space targeting small business owners. That's a lot better. Then you wanna make sure that you're giving it context.
03:30So your situation, what you've tried, what's working, what isn't working. The more specific you are, the better the output is. And I mean, it gets so much better.
03:39So now let's get on to number three, and that is the output shape. So the format, the length, the tone that you are going for.
03:46If you want a table, then say that you want a table. Don't just expect that it's going to give you a table. If you want three options, say give it to me in three options.
03:54If you want it to sound like you without it, then paste in a writing sample and say match this. Now watch me rebuild this. Okay?
04:03So I'm gonna say, you are helping me plan content for a YouTube channel in the AI tools niche. My audience is small business owners curious about AI but not technical. I post once a week.
04:13My last three videos got between 8,022 views. The 22,000 view video was a tutorial.
04:20Give me five video ideas leaning into the tutorial format with a one line hook for each and make it a numbered list. There you go. It's the same model, the same Claude, but it's now output that you can actually use.
04:35Now if you wanna play or move over here, when you don't know what context to give it, and most of the time you really don't, just flip the script. End the prompt with before you write anything, ask me questions that you need answers to do this well.
04:52Claude will then interview you. You give it answers and then it writes. The output is almost always better than anything that you'd have prompted yourself because Klaud knows what it needs better than you do.
05:04HOOKSo now that is prompting at its finest. The ceiling breaks when you stop typing context every single time and start building it in. Now we get on to step number three and that is the memory.
05:16HOOKSo just a quick detail from what we are doing because this is the gap in every other Claude tutorial. Claude has a global memory across every conversation, not just a project, not just the skill, the whole account.
05:31HOOKIt learns who you are, what you are working on, and how you like to be talked to. Now here's the move that actually unlocks this. You've got months of context built up in another AI tool.
05:42Let's say GPT, let's say Gemini, whatever it is. You don't have to start over.
05:48In the memory settings, there's an import memory option. So Claude gives you a prompt and then you paste it into your old AI tool and it spits out a summary of everything it knows about you. Then you just copy that back into Claude and now hit add to memory and then years of context.
06:05That's all being ported in just about five minutes. The single highest leverage hour that you'll spend on your AI setup this entire year, and it's such a game changer.
06:15And almost nobody ever talks about this. Now Claude knows you. The rest of the bold assumes that's too.
06:22That takes us perfectly onto step number four, that is the artifacts. So this is what I built in the time my coffee took to brew. Okay?
06:31Working IOI calculator. There's an hourly rate.
06:34It's manual hours per week, weeks worked a year, calculates annual cost, time saved at automation, IOI on a $5 setup. It runs in the browser in just one prompt, no code whatsoever.
06:49That's an artifact. So when you ask Claude for something substantial like a tool, a calculator, a dashboard, a landing page, it doesn't reply in chat. It opens a panel and then it builds the thing.
07:00So here's the prompt that I use. Bold me an ROI calculator for business owners considering AI automation inputs hourly
07:09rate, manual hours per week, weeks per year, and the outputs must be annual cost time saved with 70% automation, annual savings, ROI on a $5,000
07:22setup, clean, simple, mobile, usable, simple.
07:26But you see, I'm going into detail. And that took me about ninety seconds, maybe even shorter, and and now I have a working calculator.
07:34And I mean, I didn't actually love the colors. So what I did was I just typed in, make the button dark and the background white. And it updated it live.
07:45Now you can just talk to it like you were giving it feedback or like you were giving feedback to a designer, I should say. Then what you're do is you're hit publish, you're gonna copy the link, and now you have a tool that can be embedded on your site. Can send this to clients or you can use it as a lead magnet.
07:59Developer would have taken so long to create this and on top of it they would have charged you so much, but here you just got it done. ArtFX also handle the file uploads. Keep that in mind.
08:09So what you can do is you can drop a PDF, a CSV, a screenshot, a 200 page contact right into the chat if you want and then just watch this. So now I'm exporting my YouTube analytics as a CSV for the last ninety days, and then we just drag it over, we drop it, and then we say analyze my performance trends and visualize the key metrics.
08:30And then Claude actually generates a full analytics dashboard inside an artifact. It gives me charts, top performers, all the patterns that I would have missed. It gives it all to me, and the same move works with screenshots, contacts, photos of a whiteboard.
08:44All you do is just drop it in, and then you ask the question, and you get the analysis. Now artifacts are insane on their own, but they are disposable. I mean, a new chat, fresh slate, Claude basically forgets everything.
08:56Now the next feature is what makes Claude actually remember who you are, and that takes us on to projects. So a project is a workspace with its own memory, its own files, its own instructions.
09:09You basically just come back to it. So for our YouTube workflow, I'm creating a project called YouTube scripting and three things happen inside this project.
09:19Okay? So that is the files, I doubt my brand guidelines, a sample script in my voice, and a one pager of my audience. And these are going to live permanently over here.
09:30Every So conversation in this project has access to them. And then the next thing that happens is instructions. This is where I tell Claude how to behave inside this workspace.
09:39I basically tell it, you are helping me write YouTube scripts. My audience is small business owners learning AI tools, my tone is direct and conversational, there's no corporate language, no bullet points in the script body, always reference the uploaded brand guidelines and voice sample before nighting.
09:56And then the next thing is conversations. So now I'm inside the project. I'm typing the same prompt that I used at minute three, and I'm saying give me five YouTube video ideas.
10:06And now take a look at the defense. It knows my niche. It knows my tone.
10:10It knows that tutorials outperform opinion pieces. Why? Because I told it that in the project files.
10:16So it's not a better prompt. It's basically a better environment. And now if you wanna pair it up, account level memory is for you.
10:26The name, level, the communication style, the project instructions are for the slice of work. Voice format deliverables, keep all of them separate and if you claim everything into account level, then Claude actually starts giving you weird hybrid responses when you switch context.
10:44Now projects only work inside one project. Now what if you want Claude to behave a certain way everywhere? That's what we are going to be talking about next, and that is skills.
10:57So here's where most people stop, and it's the most expensive mistake in your whole setup. Right now, every recurring task, proposals, LinkedIn posts, script outlines,
11:07you either retype a long prompt or you paste one that you saved somewhere else every single time, and over time it adds up to quite a lot of wasted time. Now skills breaks that ceiling. A skill file is instructions that Claude reads automatically when the task matches.
11:24So you describe the workflow once, then Claude learns it, and from that point on, you just say, write me a LinkedIn post. And now Claude already knows your format and knows your tone and knows your hook style, your CTA without you having to re explain everything. And I mean, the mechanic that makes this work is progressive disclosure.
11:44Claude only loads the skill that is relevant right now. You can have 50 skills installed and Klawd doesn't get any slower, just more capable over time.
11:54So a quick note, projects are a workspace. MCP, which is the model context protocol, is the connection layer. So it's basically how Claude talks to your other tools.
12:05Skills are the how to sitting on top of both. So now in one line, MCP is the kitchen and skills are the recipe.
12:14You don't write the skill file yourself. Claude writes it based on what you described. So I would say I want a skill that writes LinkedIn posts for me.
12:23The tone is direct and a little punchy and short sentences, no corporate language, no hashtag spam, and every post opens with a one line curiosity hook and delivers one specific insight and ends with a question or CTA in under 150
12:39words. So then Claude is gonna ask me a couple of questions, I answer it, and then it generates the skill file, and then I just scan through it, I give it the few tweaks here and there, and then I just simply hit install. So now in a new chat, would say write me a LinkedIn post about why most people get bad results from AI tools.
12:57Okay. Then just watch what happens. Now the hook is under a 150 words.
13:02It sounds just like me, and I didn't explain any of that in this prompt. The skill handled all of it for me. And when it doesn't necessarily work, then the one line debug that no one else shows you, you're about to see now.
13:16So sometimes you install a skill and Claude just for some reason ignores it. Now if it does, then here's the fix. What you need to do is type in when would you use the LinkedIn post skill and then Claude quotes the skill descriptions back to you and the second it does that, you'll see exactly
13:35what's vague. So maybe the trigger condition is too narrow, maybe the description doesn't match how you actually phrase the task. Edit just one line and then you reinstall and then it works.
13:47And that's the entire troubleshooting process right there. Now you don't have to build every skill from scratch.
13:54Anthropic ships a directory, Notion, Figma, Atlassian plus official ones for Word, Excel, PowerPoint, PDF. All you do is just click the plus, you browse, and then you install.
14:05So if you've never ever used skills, that's how I'd start. Now what you do is you pick one that matches something that you would actually do, and then you just watch it fire, and then you build your own.
14:16Every recurring task in your work is now a skill. Not a prompt that you have to retype over and over, not a doc that you have to copy from. A skill that fires automatically and produces
14:27the right output every single time. Now we have to go into step number seven and that is the custom instructions. So custom instructions are the always on briefing that every Claude conversation starts with.
14:38Now before you type a word, before Claude reads your prompt, it reads your instructions, every chat, every time. Some people will tell you not to use account level instructions because they bleed into context where they don't fit.
14:54They have a point with that. Now if you write a generic 800 word bio, then the fix isn't avoiding custom instructions. It's writing them tight enough that they apply universally.
15:05Now let's explain this in five sections. Who am I? The name, what I do, what I'm building, who I serve, two or three sentences and not a not a full massive bio.
15:14And then the second thing is how I want Claude to talk to me. So the tone, the length, the formatting. I want direct short sentences, no bullet points unless I ask,
15:23no great question, um, no em dashes. Just write it down and Claude follows it.
15:30And then what I'm working on right now, which is the current goals, the current projects, and what I'm pushing on this month, Update this every week, and it keeps Claw pointed at what actually matters. So how I work, um, when do I sleep, how I make decisions, or what I need when I'm stuck versus when I'm in a flow. Now this is cool because it changes the quality of strategic conversations.
15:53And then what to never do, the deal breakers. So my no corporate hedging, no important to note, no suggesting I hire someone when I'm asking how to do it myself, no lists when I ask for a paragraph.
16:06So the sharper this section is, the less you'll correct Claude midsection or mid conversation, I should say. And then the last thing is keep it tight.
16:14So mine are always about 200 words across all five. So if you are writing more, then step out anything that isn't load bearing.
16:23And this is a briefing. Keep that in mind. It's not a biography,
16:27so you don't need to give it so much. Cool. Now I'll give it the same prompt.
16:30So before the custom instructions, it was generic. It was formal. It was bullet pointed.
16:34Now after, it's direct. It's in my voice.
16:37It's formatted the way I want it, and it's the same model, but different briefing. So now we have to go on to step number eight, and that is the connectors and the research mode. So two more pieces and the workflow is complete.
16:49So connectors, it's the plus icons and then it's the add connectors. So this is where Claude plugs into the tool that your work actually lives in. The Gmail, the dive, the calendar,
16:59the Notion, those Slack. So I'm connecting it to my Google dive and now when I'm inside the YouTube scripting project and I say pull my channel positioning dock and use it to inform the next script, then Claude doesn't need me to upload anything. It reaches into my dive, it finds the dock, it reads it, and it applies it, and the information lives where it lives.
17:19Claude pulls it on demand whenever it needs it. So now the connectors that save the most times are the ones that are plugging into wherever your actual work happens. So for a content creator, that's Google Drive or your Google Calendar.
17:34For an agency owner, it's Gmail or even Notion. And all you have to do is that they just pick one to start. Now for the research mode, it's plus icon and then it goes over to research.
17:44So it's different from web searches because web searches gives Claude the Internet for one quick answer. Research mode is autonomous. So you give it a question and it builds a research plan.
17:55It runs multiple searches that build on each other and the cross references the sources and it comes back with a fully cited report, so five to fifteen minutes for most queries. Now for YouTube, I use it for competitive research. I would give it something like research the top five AI tutorial channels above 100 k subs, what hook formulas do they use, what video links perform, what topics overlap, synthesize it into a positioning brief, and then it comes back with citations,
18:22patterns, gaps that I could actually exploit. And it used to take about half a day of manual work on my side. Now it's about five minutes while I'm making a lunch.
18:31Then all you do is just use it for anything that needs thoroughness, not speed, competitive analysis, market research, big decisions that you need to make.
18:41And now just one little bonus for you in the model selector, there's an extended thinking toggle. So turn it on for hard problems. So such as like things like strategy,
18:51financial analysis, multi step reasoning. Claude actually takes longer,
18:56but it reasons through the problem step by step. So your result is actually better.
19:01The output is noticeably better across the board when the thinking actually matters more than the speed. So here's the payoff.
19:09If you start the bandwidth chat inside the YouTube sketching project, then custom instructions are loaded, memory is loaded, the skills is installed, and the dive is connected. So I just type in one sentence, draft a video about why most people get bad results from AI tools. Then it pulls in my channel positioning dock from the dive and it uses my standard script structure.
19:28CTANow watch how it works. Custom structure is hit. The the project loads my voice and my audience.
19:35CTAThe dive connector pulls the positioning dock, the skill triggers, and it applies my script structure. And now the output that I get is a draft that I actually film with my hook, with my pacing, my CTA preferences,
19:48CTAand it's all in my voice. Now compare that to the cold start prompt that we had earlier. Write me a content plan for my YouTube channel.
19:56CTAAnd here you can see the difference isn't even close. It's the same model, the same chord, but completely different results because the system is doing the work, not the prompt itself.
20:07CTAAnd that's the workflow rebuild. Now that's pillar one. Everything else Claude co work, Claude code,
20:14CTAit sits on top of what you just built. So if you like code, if you build internal tools, or if you automate anything technical, then my Claude code walkthrough is next.
20:24That's where Claude builds apps, websites, read your entire code base, and it ships features from a single prompt. If you want Claude actually doing work on your computer while you are in another tab, opening up apps, moving files, running multistep tasks autonomously, then my coworker walkthrough
20:42is the one for you. The templates, my custom instructions, the exact skills that I use, and the rest of my workflows are all in our free school community. The links are down in the description below.
20:52Go and check it out and also check out the video that's popping up on screen right now. Like and subscribe, and I will catch you on the next one.
— full transcript
§ 05 · For Joe
Build the environment, not a better prompt.
The system > prompt thesis
Every recurring result problem in Claude traces back to a missing environment layer — the fix is never a cleverer prompt, it's installing the right layer.
- Map your own 8 layers: model, prompts, memory, artifacts, projects, skills, instructions, connectors — audit which are actually live.
- The interview flip is the fastest win: append 'before you write, ask me questions' to any vague prompt. Zero setup, immediate improvement.
- Write custom instructions in 5 sections, 200 words max. The 'what to never do' section pays the most per word.
- Build a skill file for every recurring task — Claude writes the file; you just describe the workflow once.
- The skill debug trick ('when would you use the X skill?') belongs in your standard troubleshooting flow.
- The import-memory trick (port GPT/Gemini history into Claude in 5 min) is underused and demonstrable — strong JoeFlow demo candidate.
§ 05 · For You
Why your AI answers feel generic — and the fix.
For anyone using Claude or ChatGPT
The reason most people get weak results from AI tools isn't the model — it's that they ask short questions and expect expert answers, like it's a search engine.
- Give Claude a role, context, and output shape in every prompt — three sentences of setup saves ten minutes of editing.
- When you don't know what context to give, ask Claude to interview you: end your prompt with 'before you write, ask me what you need.'
- Custom instructions take 15 minutes once and change every conversation after that — tell it your name, what you're working on, and what you hate seeing.
- If you've been using ChatGPT or Gemini for a while, Claude can import that history — you don't have to start over.
