05:58Lead Gen Jay · Talking Head
Insane Free Gmail Hacks 98% Don't Know About (Instant Trust)
Two free, buried Gmail settings decide whether a stranger ever sees your face — and both default to hidden.
July 16thThe argument in one line.
Running a business on Claude Code at scale means treating it like a small workforce: a cheaper model executes, one expensive planner reviews the plan first, and every mistake the review catches becomes a permanent rule.
Read if. Skip if.
READ IF YOU ARE…
- You already run Claude Code for real business work and want to see a full multi-session, multi-app setup in practice.
- You're deciding how to split spend across a planning model and an execution model and want a working cost example.
- You manage marketing, email, or ad campaigns and want to see an agent plan and launch a Meta/Google campaign end to end.
- You want a concrete example of converting AI mistakes into standing rules instead of re-fixing the same bug.
SKIP IF…
- You're brand new to Claude Code — this assumes you already know terminal basics, sessions, and slash commands.
- You're looking for a feature comparison between AI models rather than a real operator's day-to-day workflow.
The full version, fast.
This is an unscripted screen recording of an eight-figure business owner's actual Claude Code sessions, not a feature demo. Three terminal windows, each split into parallel sessions, run isolated apps — email marketing, a website/ad platform, and a software product — at the same time. The stack behind it: Fable plans and orchestrates while a separate ChatGPT model executes, because execution is three to four times cheaper on tokens that way. Before any task runs, an interview command keeps asking clarifying questions until uncertainty drops below 10%, and a capped, multi-round review pass catches dozens of plan holes before any code gets written. Every bug that review finds gets converted into a standing rule so future sessions don't repeat 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 →Where the time goes.
00:00 – 02:07
01 · Inside My Real AI Workflow
Cold open pitches the video as an unscripted screen recording of a real business owner's actual Claude Code sessions, in contrast to feature-focused AI content.
02:07 – 03:23
02 · My Multi-Session Terminal Setup
Explains the three terminal windows (Inbox Insiders, Web Designer, Consulty), each isolated by project and folder, and each split into multiple parallel sessions working different features.
03:23 – 05:28
03 · Claude Code Runs My Marketing
Shows Claude Code drafting an email campaign in Kit, managing backend customer data for cancellations and mismatches, and building hyper-targeted per-industry landing pages off one reusable template.
05:28 – 06:34
04 · Fable Plans, GPT Executes
Lays out the model stack: two $200/month Claude Max plans plus one $100/month ChatGPT plan, with Fable as planner/orchestrator and GPT-5.6 Sol as the cheaper, faster execution worker.
06:34 – 07:32
05 · Free Codex Skills Package
Demos his free downloadable skills package (leadgenjay.com/skills) built around two commands — codex review and codex execute — that spawn GPT-5.6 Sol sub-agents inside Claude Code.
07:32 – 12:46
06 · Test, Improve, And Ship
Walks through his 'test improve' and 'ship' commands: Playwright opens the built feature, clicks through it, checks it against saved design rules, fixes defects, then ships one feature at a time to avoid branch collisions.
12:46 – 13:46
07 · Planning Hyper Targeted Ads
Introduces the next task — building Meta and Google ad campaigns hyper-targeted to specific client avatars — and dictates a structured goal statement to Fable, deliberately scoping landing pages out.
13:46 – 20:42
08 · The Socratic Interview Command
Fable's 'interview' skill keeps asking clarifying questions — avatar wave order, budget source, decision rules, targeting method — until uncertainty drops below 10%, offering a recommended answer each time.
20:42 – 22:36
09 · Managing Your Context Window
Explains his HUD (five-hour window, weekly limit, context percentage) and his rule of compacting or clearing a session once context climbs past 30-40%, rarely letting it reach 50%.
22:36 – 26:00
10 · My Custom Admin Dashboard
Tours his self-built leadgenjay.com/admin dashboard — ads command center, funnels, AB tests, custom live chat, sales routing, lead/call scoring, commission tracking — all built and maintained via Claude Code.
26:00 – 30:34
11 · Hermes Agent Posts To Reddit
A remote Claude/Fable agent running on a Mac Mini ('Hermes') texts him a one-click approval link before posting a Reddit reply, then runs further tests — a LinkedIn comment and DM — using Composio and Unipile.
30:34 – 34:15
12 · Composio And Unipile
Breaks down the two platforms behind the agent's actions: Composio for pre-built, developer-account-free API integrations, and Unipile for LinkedIn/Instagram browser automation with its own account-risk tradeoffs.
34:15 – 37:20
13 · Turn Bugs Into Rules
Codex Review's three-round iteration cap surfaces 23, then 15, then 8 plan holes; he has Fable convert every finding into standing documentation before wrapping the session and signing off.
Lines worth screenshotting.
- Splitting AI work between an expensive planning model and a cheaper execution model can make a session three to four times more token-efficient without losing output quality.
- A Socratic-interview command that keeps asking questions until stated unknowns drop below 10% catches ambiguity before it turns into wasted execution time.
- One review pass found 23 holes in an ad campaign plan, a second found 15, and a third found 8 — all before a single line of code was written.
- Letting a session's context window climb past 40-50% measurably degrades output quality, since every later command has to be processed against everything already sitting in context.
- Documenting every bug a review session finds as a standing rule means the same mistake never has to be caught twice.
- A text-message, one-click approval link lets a human stay in the loop on an autonomous agent's public actions without opening a terminal.
- Shipping two features to the same branch at once risks the code getting confused mid-deploy, so features get shipped one at a time even when several are ready.
- Pre-built integration platforms remove the need to register a separate developer account with every third-party API a business wants an AI agent to use.
- Automating a platform without an official API through browser automation carries real account-risk and terms-of-service exposure that has to be weighed against the time saved.
- Landing pages built as one reusable template can be re-targeted to a new industry vertical far faster than hand-building a new page from scratch each time.
- A visual regression check — open the page, click through it, screenshot it, compare against saved design rules — catches defects a code review alone would miss.
- Isolating every project into its own terminal window and folder prevents contexts from bleeding into each other when several apps are being built at once.
Split Planning From Execution, Then Codify Every Bug
AI Workforce Management
Running a business on AI works best when planning and execution use different models, every session stays scoped to one task, and every mistake gets written down as a rule before it can repeat.
02My Multi-Session Terminal Setup
- Running several isolated projects requires physically separating them into different terminal windows and folders so contexts never bleed into each other.
- Splitting one project into multiple parallel sessions lets several features get built simultaneously, as long as each session's current task is visible at a glance.
03Claude Code Runs My Marketing
- An AI system connected to a backend database can handle recurring account operations like checking mismatches and processing cancellations, not just writing code.
- Drafting and reviewing email campaigns inside a dedicated email platform before sending keeps a human checkpoint on anything that reaches a full list.
- Building landing pages as a single reusable template that gets re-targeted per audience segment scales far faster than hand-building one page per vertical.
04Fable Plans, GPT Executes
- Splitting work between an expensive planning model and a cheaper execution model can cut token costs sharply while keeping output quality high, since planning tokens are a small fraction of total usage.
- Using two subscriptions with different strengths, one for orchestration and one for high-volume execution, is more cost-efficient than paying premium rates for every step of a task.
05Free Codex Skills Package
- Giving one coding assistant the ability to call a second model as a sub-agent turns a single terminal into a small team of specialized workers.
- Publishing a lightweight, install-in-two-minutes package removes the main friction people have with adopting a new AI workflow.
06Test, Improve, And Ship
- A standing command that opens a finished feature in a browser, clicks through it, and checks it against saved design rules catches visual and functional bugs before a human ever looks.
- Shipping code changes one at a time, rather than in parallel, avoids two AI sessions colliding on the same branch and confusing each other's context.
- A single command that documents what a session did, extracts reusable lessons, and prepares the session to close removes the manual wrap-up step most people skip.
07Planning Hyper Targeted Ads
- Writing a structured goal statement, covering audience, objective, constraints, and what not to worry about, gives a planning model enough scope to produce a workable strategy in one pass.
- Explicitly separating what one session is responsible for from what another session owns prevents scope creep inside a single AI task.
08The Socratic Interview Command
- Forcing an AI planner to keep asking clarifying questions until stated uncertainty drops below a set threshold catches ambiguity before it turns into wasted execution time.
- Presenting a recommended answer alongside every open question lets a human approve quickly instead of generating the decision from scratch each time.
- Testing two strategies at once, such as broad versus interest-based targeting, is a decision an AI can propose but a human still has to greenlight.
09Managing Your Context Window
- The larger a session's context window gets, the more tokens every subsequent command burns and the worse the output tends to get, so compacting at 30-40% keeps sessions sharp.
- Letting an in-flight session with active sub-agents finish before compacting avoids interrupting work that's still running.
10My Custom Admin Dashboard
- Centralizing marketing, funnels, AB tests, live chat, and sales tracking into one custom-built dashboard lets an AI assistant reference and update all of it from a single place.
- Feeding an AI assistant a log of real customer support conversations lets it find patterns and suggest changes that reduce ticket volume, not just answer new questions.
- Migrating off a general funnel builder onto a custom site removes a layer of platform-specific quirks, but only pays off once those quirks are documented as rules the AI can check against.
11Hermes Agent Posts To Reddit
- An agent that texts for one-click approval before publishing keeps a human in the loop on public-facing actions without requiring them to open a terminal.
- Running visual regression checks on newly shipped pages before declaring them live catches defects a code review alone would miss.
12Composio And Unipile
- Pre-built integration platforms remove the need to register a developer account with every third-party API a business wants an agent to touch.
- Automating actions on platforms without an official API through browser-based automation carries real account-risk and terms-of-service exposure that has to be weighed against the time saved.
13Turn Bugs Into Rules
- Capping an automated review loop at a fixed number of rounds, rather than letting it iterate indefinitely, controls cost while still catching the majority of issues in the first pass.
- After a review session finds bugs, explicitly asking the AI to convert those findings into standing documentation means future sessions inherit the fix instead of repeating the mistake.
Terms worth knowing.
- Fable
- The AI planning and orchestration layer (Fable 5) used as the 'mastermind' that plans and reviews work before a cheaper model executes it.
- GPT-5.6 Sol
- The ChatGPT model used as the token-efficient execution worker that carries out plans Fable creates, at roughly a third to a quarter of the token cost.
- Codex CLI
- OpenAI's command-line tool that lets a Claude Code session call GPT-5.6 Sol as a sub-agent to help plan or execute tasks.
- Codex Review
- A command that spawns a model to audit a plan for holes and gaps before any code is written, capped at a fixed number of iteration rounds.
- Codex Execute
- A command that spawns worker sub-agents to carry out a plan that has already passed review.
- Interview command (Socratic interview)
- A skill that keeps asking clarifying questions about a stated goal until unknowns drop below a set threshold, offering a recommended answer for each question.
- Context window
- The running memory an AI session holds. The fuller it gets, the more tokens every command burns and the more output quality tends to degrade.
- Composio
- A free platform that provides pre-built API integrations (Reddit, Zoom, QuickBooks, and others) so an agent can use third-party software without a developer account for each one.
- Unipile
- A paid platform that automates LinkedIn and Instagram through internal APIs and browser automation, used to run an agent's social actions on platforms without an accessible official API.
- Hermes agent
- The operator's personal always-on AI agent, run on a separate machine, that texts him for one-click approval before taking public actions like posting to Reddit.
- Avatar landing pages
- Hyper-targeted versions of one offer page, each rewritten for a specific industry or buyer persona instead of a single generic page shown to everyone.
- Test improve
- A standing command that opens a built feature in a browser via Playwright, clicks through it, screenshots it, and checks it against saved design rules before shipping.
- CDSP
- A hotkey shortcut for starting a new Claude Code session in 'dangerously skip permissions' (bypass) mode.
Things they pointed at.
04:30toolGoHighLevel
05:45toolFable 5
06:00toolGPT-5.6 Sol (ChatGPT)
11:45toolRectangle
Lines you could clip.
05:45
“I really want Fable to be my mastermind.”
13:20
“Fable only needs about medium effort to be twice as good as Opus.”
25:50
“I can almost guarantee that it gets it right on the first try.”
21:10
“The higher your context window gets, the more tokens you're consuming on every command and the worse it's gonna perform.”
29:10
“You have to have a special type of brain in order to do what I do.”
34:30
“23 holes in round one, fifteen in round two, eight in round three.”
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.
metaphorstory
Alright. What is going on, nerds? Figured I would make this video so you can see what actual business owners use AI for.
You're probably used to watching all of these YouTubers talk about different features and are saying that Chatchipiti just crushed cloud or vice versa. Well, you probably don't get a chance to watch what actual business owners are actually building with AI. So today, wanted to give you an inside look inside my actual AI sessions and show you all of the things that I'm actually using AI to build and all of the shortcuts, all of the tricks that I'm using to make myself myself more efficient, save credits, and actually build things that make me money.
I am not your average AI content creator or YouTuber. I'm actually building real businesses, and this is how I do it. So I'm gonna give you an inside look inside of my computer.
This is the first time I've ever done any video like this, So please let me know if you enjoy this format. It's very informal. I don't have a script and I don't have an outlet.
I'm just going to show you what I'm actually doing in my day to day. So let's go inside of my computer and let's get started. Alright.
So a couple things you need to know as you're looking at my screen right now. One, I've got a gigantic monitor where there's just no way to show you the entire thing.
So right now you can see about the right half. I usually keep all of my Claude code sessions on the right half of this monitor. And the second thing you need to know is this is not the Claude app.
We are not in the web app. So if you're brand new to AI and looking at this scares you, then you should probably skip this video, go back and watch my introduction to Claude code. It's a full one hour beginner's video.
But if you're deep in the weeds like me, I think you are gonna get a kick out of this. Now, hopefully, there's nothing sensitive on the screen. If there is, I'm gonna ask my video team to blur some things.
But this is literally just a snapshot of what I'm doing right now. I'm gonna walk you through all of the different things that I'm building right now with Cloud Code, how I'm using the different models to save credits, some of my shortcuts, and how I organize my thoughts and actually think through using AI.
And I promise you're gonna get some ideas, and hopefully, will make you better. Also, I would love your help. So this is me asking for your help.
If you're a nerd and you're deep in the weeds like me and you see a way that I can optimize or I can be more efficient and I can be more better with my AI use, please leave it in the comments. I wanna know. I also wanna get better.
Now let me show you what I'm working on. So the first thing you might notice is we've got three different terminal sessions open. Each of these terminals is split into multiple sessions.
So why is this important? Each of these, inbox insiders, web designer, consultee, these are all essentially different apps.
And I keep them isolated in different terminal sessions, different projects, different folders on my computer. These are completely isolated softwares, different websites that you log in and go into. They all serve different purposes and do different things.
This is my mailbox app. My web designer, this is actually like my leadgenj.com, which also serves as a lot of other purposes such as sales tracking and all of my websites and AB testing.
And then down here is consulting, my software products. So obviously, I've got a lot of sessions here. The other thing that you'll notice is each project, each of these terminal windows is split into multiple different sessions.
So these are all working on different features in one app all at the same time. And then you'll look at each of these sessions and you can see on the top here what it's actually working on. So fixed credit bypass limit, I'm actually adding some guard rails, widget updates, feed rebuild.
So I'm actually rebuilding our support widget so it looks more like intercom and has more functionality, refreshing our roadmap, and then refreshing our onboarding and intake process. Up here in the inbox insiders, I'm doing a public updates and change log.
And where we're stuck here now, it's actually using kit.com to do our email. So let me open up my window here. It actually added all of the contacts to this kit campaign, and it drafted this email inside of kit.
I'm reviewing it. I haven't sent it yet. We use AI.
We use Cloud Code to manage all of our email marketing, both in kit and in go high level, but we use kit because it's got stronger deliverability with our IP addresses. Now we're not only using Cloud Code to build features, we're also using it as a copilot to manage lots of different things going on inside of our business, which is what you see here.
Cloud Code has access to all of our back end database. So it knows all of our customers, their domains, their mailboxes. So a lot of times, I use it to check if there's any mismatches, any cancellations, and process those.
And I do that all within Cloud Code. Now over here in our web designer, we're actually building hyper targeted landing pages per industry.
Let me show you what that looks like. So one of my big marketing pushes right now is instead of just using a single landing page to send to all the industries, this one is just gonna be for agencies. So your agency needs an automated cold outbound system.
It's really easy for me now with my Cloud Code to spin up hundreds of these landing pages that are the same offer, but they're hyper targeted to a specific vertical. So that's what I'm doing in that session there, and I'm just refining the design, refining the copy. And if I refine it once across one vertical, now it can apply it to all of the other verticals.
I'm creating rules. I'm creating repeatable commands. And in this session here, I'm creating a reusable command so I can check logs of all of my members and see what they're actually doing.
Now this one is done, so I'm gonna go ahead and hit command end. And what that's gonna do is it's a command that I actually built when I'm done with this session. I want it to document what it did during that session.
It'll store everything to the local documents. It'll go through the session and see if there's any lessons or rules that it can actually use to update the docs to improve itself next time. And it'll finally prepare the session for me to close it out.
I like to run that before I close out any new session. Now let's just look at Consulti for a second. Specifically, I'm gonna go ahead and zoom in on this.
So let's start with this first one. It just finished a long session. As you can see, says Soul built in.
Now let me explain some of the commands that I use because I really want Fable to be my . Fable's gonna be my planner and my orchestrator. However, Fable credits cost a lot.
So instead of using Fable to actually execute the plan, I'm not using Opus, I'm not using Sonnet. I actually have GPT 5.6 SOL as the workers to execute the plan from Fable.
And that's what I probably recommend most people do. Now in order for me to pay for all of that, I actually have two Claude Max subscriptions that are $200 per month each. That's my main driver.
I'm using Claude code for most of these things. I also have one $100 per month ChatGeeBT subscription, and I use that with Sole 5.6.
This is a super powerful model. It's unbelievably good at both helping to plan and making sure the plan is really rock solid. So Fable and SOL five six will go back and forth and actually review the plan, and then I use SOL 5.6 as the worker to execute.
And it's like three or four times more efficient with tokens, and ChatGPT is pretty generous. So I'm able to get away with just that one $100 plan, and it's been really, really helpful so far.
Now here's my skills stack, leadgenj.com/skills. So let's actually go to codex, and I've got this free package that you can download that actually has all of the the skills that I'm talking about. And it's really just two.
It's codex review, codex execute. And this is going to spawn sole 5.6 agents to go through review the plan and to spawn 5.6 sole workers to actually execute the plan.
So the first thing this is gonna do is install that codex CLI. It's gonna give you a prompt to open up codex and log in to your ChatGPT. It takes two minutes.
And now all of a sudden, you can actually call ChatGPT 5.6 from inside of your cloud code session. So if you want this, legionday.com/skills. This one is public and it's free.
I'm gonna go ahead and let it push that update live. Now I do have other commands that I like to use. So for example, we've got this one here.
Commit and open the PR. This one's actually done building that intercom style widget in my claw design skin. So I wanna kinda see what this looks like.
Now I know it already did a test and prove session. So another one of my commands that I really like to use over and over again is called test improve. And what test improve is gonna do is anytime I'm building a feature or a page or something visual, I don't just want it to ship it.
I don't just wanna trust it that it looks good and it acts good. I want it to actually open it up in Playwright, click through it, take screenshots, and make sure that it works the way that it's supposed to work and looks the way that it's supposed to look. I also have all of these design rules baked into my my projects so it knows, okay, this headline is too big or it doesn't render properly on mobile to go back and fix it and iterate and only show me the final design once it's actually complete.
I already saw it go through that, so I'm gonna go ahead and say slash ship, and now it's going to actually push these changes live. It's going to update the feature documentation, and it's gonna do several other really important checks, like a codex code review to make sure nothing is leaking.
There's a lot of different things that go into that ship command. I actually built this with Fable and rebuilt it with Fable. So it's perfectly optimized so the code is deployed the right way, especially since right now we've actually got two features pushing live, which I don't usually like to do at the same time because right now we're all all on the main branch and it could overlap with each other and the code could get confused.
So I like to ship one at a time. But because of how dialed in this command is, it's probably going to do the right thing. Switch the branch, open a PR, all that boring stuff that I don't even fully understand but Fable does, Now it's doing this in the correct way.
Now you might also notice that I'm doing these little hot keys and these terminal windows are like snapping to different locations on my computer. There's actually an app called Rectangle. It's it's free to download.
And now I can just like hit different keys and it goes to different locations on my computer. Makes it really easy if you're managing a lot of different windows. Alright.
Now what I wanna do is open up a brand new terminal window and start a new project or task from scratch, ideally a hard one. And then if you stick around for just a little bit, I wanna show you something else really cool that I'm working on. If you check this out right here, it might be a little small for you to see, so let's make it bigger.
This is actually me remoting into my Mac Mini, and I've got another Fable session working over here. Let's close that out. And what I'm doing over here is actually I have a Hermes agent running on this Mac Mini.
And I'm using Fable to coordinate and train and reiterate and build this Hermes agent. And right now, I'm actually working on having this Hermes agent manage my school, my YouTube, my Instagram, my Reddit, and my Linked It's gonna have access to all of those platforms. It it already does.
And I'm gonna show you exactly what I'm training it to do and how I'm training it after I launch this new project. So actually, let's go ahead and get this thing started on the next test. Continue to the next test.
See, it hit about a 50% context window, and I like to keep it ideally under 30. Sometimes I let it balloon to 40, very rarely up to 50. But after that, I'm almost always compacting and starting with that new test.
So that's what happened here. It wanted continue on to that Reddit test, so it wanted to find a post inside of my subreddit community and draft a comment. So my Hermes agent is going to be able to post in my subreddit, respond to comments inside of my subreddit, and do some really awesome things.
And right now, Fable is building building out the entire toolkit and skill set to do it. So I'm gonna go ahead and say continue to the next test, and Fable five is going to read this and build that next test and continue. So I'm gonna go ahead and minimize this, And then let's open up a brand new terminal window and start off with a new project.
So I'm gonna come into file, open, and I'm gonna open my social media tool. I actually have this configured in a really annoying way where I have to come down here, go back to my web designer, and then then open the social media tool.
So let's go ahead and make this bigger, and we're gonna start a brand new project, a brand new session. I'm gonna zoom in and make this easier for you to see. When I start a new Claude code session, I have a hotkey CDSP.
That means Claude dangerously skipped permission. So it's gonna open it up in bypass permissions mode. Now as you can see, we're on Opus 4.8 with extra high effort.
Now I wanna switch to Fable for this next task. I'm gonna go to model, and I'm gonna switch to Fable.
Now I don't wanna use Fable on extra high effort because that is gonna burn tokens like crazy. It does not need that. Fable only needs about medium effort to be twice as good as Opus.
So now we're on Fable, we're on medium effort, we're ready to start the next project. Now this social media tool project is the one connected to all my ad accounts. So what I wanna do now that I've got all of these avatar pages.
So what are these these avatar pages? They're the ones that I just showed you. These avatar pages are hyper targeting specific industries, specific niches for my offers.
That's gonna make my marketing a lot more effective because now I'm gonna be able to build audiences, build copy, build graphics around that specific avatar. When somebody thinks you're talking directly to them, they're more likely to buy.
But in the past, it was really hard to build all of these different pages at scale, test all of these different pages at scale, build all of these different ad creatives, and ad copy, and manage it, and launch it. However, now that I've got Claude connected to my ads, it can make ad creatives. It can launch all of those ad creatives on Meta, Google, OpenAI, and it can deploy all of these new pages.
There's no excuse for me not to be doing this. So that's what I'm working on now. We just finished the landing page setup of that, and now I'm going to ask the social media tool to do the ad side of this.
So I'm gonna go ahead and prompt it at what I want it to do. Now Fable's very good at thinking all the way through a task and doing it right the first time. So I'm gonna say slash goal.
This is only something that I use with Fable. And then I'm gonna tell it exactly what I wanna do, what I want what I want to achieve. Alright.
So let me read you what I just wrote to Fable, and I'm gonna explain my reasoning for all of this. Goal. I wanna build ad campaigns that are hyper targeted to specific client avatars, such as agency owners, cybersecurity, staffing, SaaS companies for our 97 cold email offer.
We can build the custom landing pages separately. I don't want it to worry about that. That's gonna be done on my web designer app.
Please review our top client avatars, and let's plan the meta and Google strategy for top of funnel ad campaigns that are hyper targeted with the ad copy and creative. Review our creative and plan to generate new hyper targeted creative using GPT image two edit to remix top performing creative. Do the full plan and use codex review to help.
This means it's going to call GPT Soul 5.6 to actually make sure that the plan is solid and to interview me to clear up any unknowns. Now this one's pretty important too. I have an orchestration layer called Oh My Claude.
This is the one that I prefer. And it comes with a couple really important skills and commands and plug ins.
One is this interview command. Now what this interview does is it will assess how much clarity it has on the goal that I'm giving it.
And if there's more than 10% unknowns, it's actually going to come up with questions and keep asking me questions until it clears up all of those unknowns so that it can do the very best that it can on the actual execution. So I always try and think, like, it is it gonna do something I don't want it to do? Is it gonna be effective with credits?
Is there anything I'm forgetting or should be including in this? Is there anything I'm telling it that it won't understand? I already know it has access to my client's avatars.
I already know it knows what GPT image two is, and it knows what remixing is. So I'm gonna go ahead and let that ride. Now my expectation is this is going to take a while.
I just got that started, and this is probably gonna be ten minutes before it has any questions for me. Main reason being is it's going through all of our meta ad strategy, Google ad strategy. It's gotta go talk to Codex and make sure that the plan is solid.
So this is doing a whole lot of stuff. So I'm gonna go ahead and minimize this and let it do its thing. Alright.
Now while that session runs and that one's gonna take a while because we're still in, a deep planning session, let's come and check on the progress of all of my other ones. Now there's another way that I organize these. I don't necessarily need to see all all of these at once.
I only really need to see the one that's active. So in all of these sessions, this is the only one that I need to take immediate action on. So I'm gonna make sure that this one's up front, and I just drag those other ones into the same window.
I'm waiting on that email campaign. Coming into here, I wanted to check if those pages are officially live. I think they are.
I'm gonna go ahead and tell Claud it's doing a good job. And then this one, I ran end, and it's now done. So I'm gonna go ahead and close that one out.
Down here, it looks like this update is in the process of shipping, and this one still needs me to merge a couple of these different updates. So what I'm gonna say here is wait for other updates to finish shipping, and then merge your fixes in your recommended order to go live.
So now it's gonna go ahead and monitor that other code that's getting shipped, uh, and it's gonna start a worker to keep checking. So if you don't wanna always just check and, like, you wanna wait for one code to finish for another one, I like to actually just tell it to wait for the other code to finish and then go ahead and run ship.
This one here looks like it's done. It shipped the important code. Everything else is not urgent.
I'm gonna go ahead and close that one out. Now down here, it actually did find a little bug where we've got an old televideo everywhere that was not properly replaced with our correct video. However, it was on, like, old homepages, so it's not super important.
I'm just gonna say leave them for now and end. And on the web designer, everything's set. The spacing fix is live.
So I'm gonna go ahead and say, test improve to make sure spacing fix is perfect, and then ship. So it's gonna actually check that its design changes were good.
It's gonna continue to iterate until it's good, and then it will ship. And my social media tool is done. Ready for the Socratic interview.
So I'm reading this as five top level components. Avatar selection. Yes.
Meta top of funnel build. Yes. Google top of funnel build.
Yes. Creative generation. Yes.
Landing pages. Deferred. Perfect.
Looks right. It's essentially splitting this into the five phases and five things that it needs to do. So it's using this interview skill right now to clear up any unknowns.
Running the interview to clear unknowns. So next one, which avatars go in wave one? See, data led, j s four.
Alright. Let's go with the data led. I usually just go with the recommended, but sometimes I don't.
So I always like to like it to do this anyway and always present me with its recommended solution. And if it's Fable, its recommended solution is likely going to be right. Now, you might also be seeing some scary numbers down here.
Let's answer this one, and I'll talk about those numbers. Where does the budget for the three avatar campaigns come from? I'm gonna say reallocate the budget.
That is its recommended solution. And now let's talk about these while it's coming up with the questions. So this is your command line interface or your HUD as Oh My Cloud likes to call it.
I like the things to be here that I wanna see that are important to me day to day. For example, what sub agents are running? How much credit usage have I gone through?
Now let's answer this next question. What's the decision rule for each avatar cell? Your existing doctrine, fair chance gate, no pause under.
So these are all of, like, the rules that I have for running meta ads, when to turn them off, when to lean in, when to increase budget. All of these different skill sets that make up the AI system. So let me just take a little break right here and talk to you about these AI systems.
So let's actually bring my browser over here, and let's talk about systems. Lead Gen j's AI systems. Now if you want to install any of these systems, they're all included in my AI automation insiders.
So instead of you just figuring out how to use Cloud Code to run your meta ads the right way so that you're actually profitable and not burning money, you'll download my Meta Ads Pro system. It is not just a skill on here's how to launch Meta Ads. It actually has all of the rules built in, all of the gotchas, all of the mistakes that could possibly be made that I've made myself and other people have made all baked in.
So that when we're launching new campaigns or analyzing campaigns and deciding whether to add budget, remove budget, this is an underperformer, or it's time to cut this creative because it's not working. It will screw up if you don't have all of the right, not just rules in place, but way of gathering information and processing information.
You don't have conversion tracking set up. Right? If you don't have Hiros or something else to compare it to, you're gonna make the wrong decision.
So all of these AI systems are available to you as part of the AI automation insiders. Highly recommend joining if you haven't already. And I really like its recommendation here.
The decision rule is I want it to beat the generic. I want it to beat the control, which is me just targeting everybody. If I'm doing top of funnel and I'm targeting a specific niche with specific messaging, it should outperform a very broad advertising campaign.
So it's going through here. Now, what do I have in my hub? I have my five hour context window from ClaudeCode, which is 63% right now with two hours remaining.
I've got my weekly. We'll come back to that in a second. Let's see.
How should avatar hyper targeting actually work on Meta? Now this is awesome because I've never run hyper targeted ad campaigns with Cloud Code yet, and it already knows the right answer, which is that it should use the copy and creative as the targeting.
And it's giving me some options here that I'm actually considering a lot more than just going with the recommended, using the creative as targeting, which I'm pretty sure is the right thing to do. The other option is interest act.
So if they're agency owners, check that interest in Facebook to say agency owner. I would love to actually AB test both.
So I'm gonna I'm gonna go ahead and do that. I'm gonna say AB both. And that way, it's gonna build two separate audience types for each one.
One with broad targeting and let the copy and creative drive meta to put it in front of the right person, and the other with those interests stacked up. K. Let's come back to the HUD.
We've got my five hour window, my weekly limit, and my current context window. Alright. This is distracting, so let me just finish telling you about my HUD first.
Let's look at this one. My current context window out of a 100%.
Now Cloud Code has a context window of 1,000,000 tokens, but I never like it to get that high. The higher your context window gets, the more tokens you're consuming on every command and the worse it's gonna perform on every command.
Because if the context window's 29%, it's now reading through the entire 29% before it does the next thing. So if it doesn't need to do that, I like to compact or clear.
Now because this one's in flight, I can see that there's two active sub agents here that are both doing something. I'm not gonna compact or clear. I'm gonna escape, let this thing do its thing.
But when my sessions do get to thirty, forty percent, I almost always wanna compact and continue from there. It's gonna make it a lot more efficient, and it's gonna save you a lot more tokens. Alright.
Let's come back into my interview. How should avatars be targeted on Google?
It wants to try Pmax or a new search campaign. I actually want to see if it can do both. Let's try both search and p max.
So I'm gonna really on this. It's gonna be doing a lot of stuff for me. And if you're interested in seeing the result of this, seeing how I've set up all of my Google rules, my Google preferences, there's a whole AI system, you guessed it, for running Google Ads and OpenAI Ads that does all of this thing.
It'll help you with tracking. It'll help you set up all of the code. It'll help you actually build the creatives, deploy the creatives.
It'll check if something's running efficiently, if you should cut budget on something. If something's not working, why? So skip all of the pain and just use my systems.
They work really well. Oh, this is gonna be fun. So it's it's about to post on my subreddit from my remote agent.
This is gonna be very cool if you wanna see it. But for now, I wanna finish my interview. Until the custom avatar landing pages are built, where do the avatars send traffic to?
Wait until landing pages. Good questions. All around.
It's really filling in a lot of context. Obviously, don't want it to just, like, deploy all of the ads with no landing pages. Those landing pages are actually ready.
Uh, let's actually show you how how cool this is. Now this is my custom built, like, website dashboard, lead gen j admin. So if you go to leadgenj.com/admin, you obviously can't log in, but this is what it looks like.
It's got all of my marketing, so you can see my ads command center, all of the things going on on Meta, all of my change logs, all of my rules. It's got all of my website pages, so all of my funnels. This is where the avatar pages are landing.
My offers, tools, lead magnets. All of my AB tests live here as well. Now I am constantly running AB tests, and I've got an entire skill that literally just goes through the AB tests, sees what's working, checks previous tests, helps me come up with the next one, reviews the heat map and user session recordings.
And then we're just always testing different different strategies and different things here. We built our own live chat. So if you go to any of our pages and interact with our live chat, it's completely by me.
That way, when I come into Cloud Code, I can just say go through all of our previous conversations, and let's see if we can optimize the chat and cover more ground. That will decrease support tickets and increase conversions.
There's just a ton of really cool stuff on here. So if you ever do wanna migrate from, like, ClickFunnels or Go HighLevel or WordPress onto your own custom site, there is a ton of learning curves that you're gonna have to get through.
I can't even tell you how many sales I probably lost with little annoying just like gotchas. Like, oh, who knew, like, Cloudflare worked that way? Or who knew it would render my checkout page like that?
Or people would bounce for this reason. So all of those rules and gotchas are built into my AI systems. So if you do wanna migrate over here, highly recommend it.
It's also integrated with all of our sales tracking. So all of our forms, all of our sales routing forms, it's all integrated here. So I'm tracking our entire journey, and we've got full sales intelligence.
So I've got lead scores. I've got call scores. I've got a sales director evaluating all of our sales team's calls.
We've got commission tracking and payouts all built in here, all automated. Fully up to date dashboard with reporting for sales and our marketing team. There's just no way to do this all with any tool that's on the market that I've ever seen.
So you got all these data sources all coming in. We've got GoHighLevel. We've got WAP.
We've got Stripe. We've got Hiros, and they all need to communicate and be accurate. The only way to coalesce them nicely and have Cloud Code help you do it is by building your own custom sales tracking, sales intelligence dashboard.
So, yeah, guys, if you want to build something like this for yourself, including, like, sales routing forms that scores the lead and puts the best calls on your best closers calendar, does all of the research and sends your sales team of, like, a full research document before and after the call, after the call is scored.
All custom, all here, all available to you as part of the AI automation insiders. It finally reached a confidence level that it was comfortable with, and now it's launching Codex Review. So it's using Codex to actually review the plan, find holes in the plan before we actually execute it.
And when we do finally execute this plan, it's gonna say, you know, ready, found this. I'm gonna do codex, execute when ready.
Now I would usually wait for the plan to be done before showing you that or before putting it in there. But I'm confident enough because we did the interview process, because it's planning and orchestrating with Fable, that when it does read that, it's gonna hold until the plan is ready.
And then it's gonna use Codex Execute, which spawns Soul 5.6 sub agents to actually write the code, to carry out everything that we just planned. And it's gonna use Fable to orchestrate and mastermind all of that.
This is gonna save me the most credits, and it's gonna be the best possible work on the first try. I can almost guarantee that it gets it right on the first try. Alright.
Now, let's have some fun. I'm gonna open this up so we can see it a little bit better. So it looks like it sent me a text message with this approval ready.
So let's go ahead and check. This is my Hermes remote agent here. It's Bob.
And one social draft waiting for my approval. So let's go ahead and click it and see what happens. It's gonna open up this link that it generated for me to one click approve for the post.
Reddit, comment, reply, urgent. It's 85% confident.
You nailed the pattern. I'm just gonna go ahead and approve it so that it try it sends it. So approval is done.
Let's come back in, and I'm gonna say approved it. Send the reply. And let's see if my Hermes agent is successful.
Uh, you probably just saw a whole bunch of browser open on localhost. This is likely my other Cloud Code session running test improve on these pages. So this is the recruiting page.
Look at this. Your staffing firm needs an automated cold outbound system. So it's actually gonna go through and run those tests on these pages to make sure that they're solid before actually shipping the pages live.
So love that it's doing this. Let's come back over here. This is the beauty of running all of these Cloud Code sessions at the same time.
You get so much done. You do have to stay focused. You have to have a special type of brain in order to do what I do.
Uh, now let's see what it's doing here. Comment posted. Calling Composio.
So I promised you some sauce. My Hermes agent is actually using a platform called Composio to orchestrate a lot of the APIs that are more difficult.
Now while it's running, let me just talk to you a little bit about Composio and AI agents. If you don't have an agent that you can text, you probably need one. I'm actually working on an offer to set them up for my clients, especially AI automation insiders with, like, a single click where you can just go through and integrate it and set it up in literally, like, two minutes with no background info.
If you do have a little bit of experience, use Fable to go through all of your projects, all of your preferences, all of your skills and rules, and build you a Hermes agent. It's not that hard.
Fable's gonna do a really good job of it. And, yes, I do have an AI system that's gonna help you deploy one really quickly, really easily with all of my rules and gotchas and skills that you're probably gonna wanna use in your own business, like managing your ad accounts, managing your social media accounts, making graphics and images for you, making scripts for you.
My Hermes agent is trained to do all of those things, and yours will come pre trained if you join AI automation insiders and get this AI system. Just like that, test two. Reddit is passed.
Let's go. It was able to post to my subreddit and reply to a comment on my subreddit. That is awesome.
And test three, LinkedIn. Comment on Jay's own latest post plus a DM to Justin or Madison via UniPi. Now that is two softwares that you now know about that it is using to execute all of these actions.
And because you're still here and because this is a very unscripted video, now you've already become aware of two very important softwares that my agent is using in order to do things that your agent probably can't, and that is Composeo and Unipile. So let's talk about those.
I'm gonna go ahead and hit start test three right now, so it goes ahead and get started. And then let's talk about Composeo and Unipile. This is still forming a plan.
Awesome. It hasn't even read Codex Execute yet, so that's what I wanted. Good.
And now let's open back up my browser. Also, you'll probably see, you know, since this is unscripted, I'm just sharing stuff. This Claude design session.
Anytime I'm doing some kind of new design, new template, for example, I showed you earlier that I was having Consultee build a new support widget intercom style. I didn't just trust Fable or Claude Code to build the design for that widget well.
So I wanted to build the widget here in Claude Design. Here, it's much easier for me to go back and forth with it, suggest specific edits, specific changes, and it does a much better job understanding UX and UI and design and mobile rendering and desktop rendering.
So anytime I'm building any kind of new design, I'll do it here in Cloud Design. And as you can see, we've got this fully interactive support widget. This is what I'm installing in consulting right now.
Probably looks familiar because it's based off of Intercom. We've got this home section, help section, AI messages, updates, so you can see all the new stuff we're shipping, all of the tasks remaining for your onboarding, and we've got tweaks.
So this is what the close state looks like. And when you click on this, it opens straight up to updates, and then you close it, and it's gone. So all of you even those, like, little interactions, little tweaks I built here in Claude design, and then I'll just literally just grab the project HTML, give it to Claude code, and say use this design.
And it will, and it'll be perfect the first time. So if you're doing any web design or widget design or app design, this has been the workflow that works best for me. So that first software that you saw my AI agent using is called Composeo.
Composeo is a free tool, completely free to use, and you can use this to integrate pretty much everything. Now what you'll notice, especially if you have an AI agent or a Claude code and you wanna, like, get your Claude code to use your QuickBooks or get it to use your Zoom, is it's a pain in the ass. You have to make a developer account on each of those platforms.
Not a lot of fun. However, with Composeo, they have integrations for all of those things.
So you literally sign up for a free account on Composeo. You'll go to connect apps, and they have prebuilt apps for each of these, and you can just come through and integrate pretty much all of your software. There's almost nothing that it doesn't have.
There's some weird ones it doesn't have, like Hiros, but all of the ones that you probably wanna connect that are a pain in the butt to connect, like Zoom, like Reddit. Reddit is a perfect example because it's what I'm using for this AI agent to post to my subreddit, to comment on my subreddit. It's doing that through Composeo.
Now, in order for me to have done this without Composeo, I would need an approved Reddit developer app. That is just not something I'm willing to go through in order to comment on my subreddits. I would just rather have a human do it.
However, Composeo has already prebuilt that app, and all you have to do is connect everything on Composeo, give your agent or your Cloud Code access to Composio, which is super easy.
They've got a whole little install page that just walks you through the install for AI agents or for Cloud Code, and it'll have all of the tooling and the MCPs. So not all of these apps are easy to use with their API. They're in charge of keeping track of the tooling, making sure that the documentation is easy to read.
So when it's using one of these apps through Composeo, it's probably gonna do a good job. So Composeo, highly recommend.
Everybody who has an AI agent should be giving it access to your Composeo io so that it can access all of these apps without you having to manage the integrations for all of them. And the second software that you might have seen is called Unipile.
This one's a little bit more black hatty. Now Unipile, as you can see here, we've got my LinkedIn and my Instagram accounts logged in and running. And now it can do things using my LinkedIn and my Instagram that would otherwise be impossible without really advanced browser automation.
Unipile manages the internal APIs, browser automations, prox ies to keep you safe, to keep your account safe.
Now, no promises. I'm not saying go use this and you'll be safe because it's probably against the terms of service for Instagram and LinkedIn, but it is a very popular tool.
A lot of people use Unipile. It is very affordable. And now you can do all of these really cool things with these platforms like Instagram, like LinkedIn, and build entire apps off of it.
So if you wanna start a business building like a LinkedIn reply agent, we're actually doing that at Consulty right now using the Unipile API. So this is still in beta, but if you come into Consulty dash LinkedIn inbox, you can actually sync your LinkedIn account and start building AI reply agents for your LinkedIn account.
But once this goes live, you'll be able to come in here, sync your LinkedIn account, and manage your LinkedIn inbox, build AI reply agents for LinkedIn all here in consulting. And we didn't build those LinkedIn integrations and browser automation.
I that was not worth the lift for me. We're just using Composio as the back end to manage all of the integrations, manage the connections, and manage all of the actions and features that it can actually do.
So you build an AI agent, you give it access to Composeo and Unipile, and now it can do a lot of really, really cool things. Alright. And now I'm gonna wrap up this video.
So let's just check all of my sessions, see if there's anything new to do, and then I will let you be on your way. And if you found this video valuable and you've learned anything, even one thing so far, please drop a comment. Let me know what you learned.
Let me know if there's anything I can learn, and please follow this channel if you wanna actually build things with AI and not just hear about the new feature announcements. So what we see here, Codex review hit the three round cap on iterate. I don't want it to iterate unlimited numbers of times that'll burn through all of my credits, so I gave it three rounds.
And it found a lot of holes. 23 holes in round one, fifteen in round two, eight in round three have been addressed in plan version four. This is why this planning phase is so critical.
Because if you can identify all of these bugs and gaps in the planning phase, execution's really easy. And execution can be done with cheaper agents like Sol 5.6. We're gonna do one more round.
Because it found eight in the previous round, it might find another one in the following round. Now what's really great about these review sessions, letting Fable talk to 5.6 solo and find all of these gaps and holes, is you can turn them into rules.
You can turn them into standing documentation so that future requests like this to build a new ad campaign strategy already have all of those findings and gotchas documented.
So it doesn't have to do this again. It can go check the documentation and avoid catching all of those potential downfalls.
So I'm actually gonna go ahead and hit the up arrow. I'm gonna clear my command to execute because it found so many issues.
And I'm gonna say, based on findings, gotchas, bugs caught during planning, please update our rules and documentation to improve future tasks that are similar.
So before actually executing the plan, I want it to document what it found, the bugs that it found, so that future sessions can be more productive and they don't catch as many bugs. So I should probably make a command for this because I use this all the time because it always finds problems and we fix the problems. And you wanna make sure that you never have to have the same bug twice or fix the same bug twice.
And this is how you do it. You have to document the bugs, add it to the rules so that future sessions can just read those rules and avoid the same problem.
I'm gonna go ahead and minimize this, and let's see what else we're working with here. Still waiting on that draft to send. This one finished.
Test improve. Spacing fix is correct and live. Interaction sweep is clean, so it interacts with everything.
Two defects found and fixed. Awesome. So during that live test improve session where it opened up Playwright, clicked through things, took screenshots, It actually found more defects, and it fixed them live before shipping.
And all of it is now shipped and live, and I'm going to go ahead and end here, which is going to document everything we did in the session, and that's it. Alright, my friends. We covered a lot during this video, and most of all, I enjoyed it because I actually got to do real work that I'm actually doing day to day that's distracting me from making content.
So if you like this format, please let me know. I will do more of it. This is fun because I don't have to make something up for you.
I can just show you what I'm doing and teach along the way. Drop a like if you enjoyed this. If you hated it and you want me to go back to a more scripted, structured format, let me know that too.
But either way, you should absolutely follow this channel and you should absolutely join our free AI automation community inside of Sku. We've got a free version and a paid one.
The free version still has a lot of really great stuff and freebies in there, but there's no reason not to be in there. But if you want everything, you want all the AI systems, you want the agents, you want coaching with me, you should join the paid one. Link is down below as well, and I will see you
The Hook
The bait, then the rug-pull.
Most AI content is a feature demo. This one is a screen recording of an actual operator running three isolated terminal setups, a tiered planner/executor model stack, and a text-approved autonomous agent — during a normal workday, unscripted.
Named ideas worth stealing.
05:45model
Fable Plans, GPT Executes
- Fable 5 — planner/orchestrator (Claude Max, medium effort)
- GPT-5.6 Sol — execution worker (ChatGPT, 3-4x more token-efficient)
A two-tier model stack where the expensive, high-reasoning model only plans and reviews, while a cheaper, faster model does the actual execution work.
Steal forany budget-conscious multi-agent build where planning quality matters more than execution cost
13:46concept
Interview / Socratic Clarification
Before executing a stated goal, an interview skill keeps asking clarifying questions — each with a recommended answer — until unknowns fall below 10%.
Steal forany AI task where an ambiguous brief tends to cause rework
07:32list
Test, Improve, Ship
- test improve — Playwright walkthrough, screenshot check, fix against saved design rules
- ship — merge, deploy, update feature docs, run a code review pass for leaks
A two-command pipeline that verifies a built feature actually works and looks right in a real browser before it's allowed to ship.
Steal forshipping any UI change without manually QA-ing it yourself
34:15concept
Codex Review Round Cap
A plan-review loop capped at three iteration rounds so it keeps catching issues without burning unlimited credits.
Steal forcontrolling AI review-loop cost while still catching most issues
20:42concept
Context Window Compaction Rule
Compact or clear a session once its context window passes roughly 30-40%, since a fuller context makes every later command slower and lower quality.
Steal forkeeping any long-running AI session fast and cheap
How they asked for the click.
VERBAL ASK
25:08product
“Highly recommend joining if you haven't already... you should absolutely join our free AI automation community inside of Skool. We've got a free version and a paid one.”
Pitched organically mid-workflow while explaining the ad-strategy rules baked into his paid systems, then repeated directly in the outro with a free/paid tier split and a link in the description.
MENTIONED ON CAMERA
Visual structure at a glance.
