First we build the thing that does the work. Then we point it at the market and get you found. Every step is numbered, and you build all of them today.
The strip along the top of every slide tells you which step you are on. Steps 03 and 09 end with ten minutes where you build, not watch.
The engine. GPT-5, Claude, Gemini. Swapping it is like swapping cars — they all drive.
The app that lets a model actually do things instead of just talk. Claude Code, Codex, Cowork.
Everything it knows in this one conversation. It has a ceiling, and past it, it forgets.
A box, a chat history, an upload button. Software — it can be redesigned next month and often is.
Running in a datacentre. Identical whether you reach it from the website, the phone or a desktop app.
OpenAI. The default. Biggest and most general.
Anthropic. Stronger on long documents and written rules.
Google. Already inside Gmail, Docs and Search.
Microsoft. Windows and Office, largely the same models.
The app decides what the model can reach. A website can only read what you paste into it. A desktop app can read a folder on your machine — which is the entire difference between steps 01 and 03.
It has read an enormous amount of writing, and given some words it works out what words most plausibly come next. Everything else falls out of doing that extremely well.
Nothing is being looked up. There is no filing cabinet behind it. A fabricated statistic and a real one arrive in the same tone of voice.
Names, dates, prices, statutes, quotes — anything about you. The more precise the claim, the harder you check it.
This is called a hallucination, and it is not lying — it has no idea it happened. One habit fixes it: anything stated as fact about your business, your industry or the law gets checked before it leaves the room.
Everything further up the conversation is still in front of it. This is why a long thread gets better as it goes.
Close the tab, come back tomorrow, and none of it survived. Your name, your business, your standards — gone.
So people re-explain themselves every morning, get a slightly worse answer each time, and quietly decide the tool is overrated. It is not overrated. It has amnesia — and amnesia has a fix.
A brilliant consultant. You still go and do the work.
That same consultant, hired in-house. It breaks the goal into a plan and executes it.
No clock, no drift, no waiting on Monday.
Breaks the goal into steps without being asked.
Ten jobs at once costs the same as one.
That is the whole salary.
And on day one it knows absolutely nothing about you. Onboard it the way you would a person — context, tools, skills.
OS/ ├── CLAUDE.md and AGENTS.md ├── README.md ├── context/ │ ├── about-me.md │ ├── my-values.md │ ├── my-offer.md │ ├── ideal-customer-profile.md │ ├── brand-voice.md │ ├── business-info.md │ ├── stack-and-machines.md │ └── memory.md ├── .claude/ │ ├── skills/[name]/SKILL.md │ └── agents/[role].md ├── Kabuto Club/ │ └── CLAUDE.md ├── Relvara/ ├── Zensen Media/ ├── Personal Brand/ ├── Real Estate (eXp)/ ├── Jarvis/ ├── Personal/ └── Active/
Corrections become durable rules in context/memory.md.
Each reusable workflow gets its own folder and SKILL.md.
A pillar-level CLAUDE.md stacks with the root instructions.
business-info and stack-and-machines came later, when they were needed. Active/ holds one-offs until they graduate.
Tell it where the rest of the context lives.
Read only the files this task needs.
If the answer is not there, ask — never invent.
Who you are and what you own.
Principles and non-negotiables.
What you sell and for whom.
The customer and their actual problem.
How you sound — with real examples.
Corrections worth carrying forward.
Markdown, not PDFs — it reads them straight through, which is faster and cheaper. These files are assets.
You cannot see it, edit it, or take it anywhere. Ask about a breakup on Monday; find it in your landing page on Friday.
Plain markdown in your own folder. You read every line, and any harness can pick it up tomorrow.
“Never make up examples. Use real ones or ask.”
It writes the correction back as a short, reusable rule.
It checks the rule before doing the work again.
Type slash and it lists every skill and command you have. This is how you see what your agent can already do.
Tag a file or folder and it loads straight in — instead of the agent burning a whole loop hunting for it.
Wispr Flow turns talking into clean text anywhere. The source spent three hours talking his context files into existence.
You have eight minutes in a moment. Talking is roughly four times faster than typing — and the interview goes deeper when you are not spelling it out.
Who you are, how you work, the standards that never change.
What this one pillar is for — its goals, boundaries and local rules.
Only the spines are loaded — every skill's name and description.
It pulls the book off the shelf and reads it only when the job calls for it. A hundred skills, almost no weight.
It stops being a chat window and becomes the layer that sits on top of every tool you own.
A library of scrapers. Instagram, LinkedIn, Maps, Reddit — almost anything public.
Web search finds a page. This one actually reads it, colours and buttons included.
One connection carrying all the others — and more than one Gmail.
A clean browser it can drive, so it can check its own work.
Your browser, already logged in, for the jobs that need an account.
Images and video, which the model cannot make on its own.
You never name the tool in the prompt. It looks at what it has and picks — the same way a new hire looks at the apps on their desktop.
Read what came in, draft the replies, and flag only what actually needs you.
Pull yesterday's numbers and write the one page you want to wake up to.
Keep it drafting, never sending, until you have watched it do the job right by hand at least once.
It only fires while the machine is awake — which is why people end up putting a spare Mac in the corner. Worth knowing, not worth buying today.
You ask for something. It goes and does it.
A trigger or a time sets it off, not you.
One finds the thing, another writes it up.
You stop assigning work and start approving it.
Three words. That is the whole prompt.
It can only do that if it knows what you know — and most of that was never written down anywhere. Five minutes a day, dictated into one file. Easiest first proactive job: a watchdog that stays quiet until something looks off.
Getting your business named inside the answer an AI gives back — instead of ranking as a link that somebody might click.
You showed up on a list. The person scanned it, judged for themselves, and picked. Being on the list was the whole job.
There is no list to scan. It reads everything, decides for them, and answers. If it does not say your name, you were never an option.
It also gets sold to you as GEO, AI SEO, LLMO or generative search. Same thing, four price tags. What it actually asks of you is the rest of this hour.
Sometimes three, sometimes ten. Six is the average.
Not ranked low. Not on page two. Absent.
Not one name holds even a 15% share. The seat is open.
If you are not one of the six, you did not lose to them. You were never in the room to lose.
The old game is shrinking, not gone. The new one is already here — and it moves in days, not the years SEO took.
of what ChatGPT cites when explaining why it recommends an agent comes from two places.
Fill in every single field. Leave nothing blank.
Which is what we spend the rest of this session on.
You do not need page one of Google. ChatGPT pulls from position 21 and beyond almost 90% of the time.
Nobody can sell you that. The answer changes with who is asking, their history, and their exact words.
Raise the odds you are one of the names it gives — to the customers you can actually help.
It reads your page.
It files the page away so it can be found again.
It decides where you show up when someone asks.
Every AI answer sits on top of those three. If it cannot crawl you, none of the next six slides can help you.
Answers when you are the right fit.
Answers why your service is different.
Answers where you are relevant.
Answers compare and prefer.
Length barely moves citation. Speed to the answer does. Never bury it in the middle.
Who is the best realtor for first-time buyers in Cincinnati?
Real questions, services, areas and proof — from calls and emails, not from your imagination.
One original answer per question. Its own URL.
Check every fact. Read every comparison twice.
Update the useful pages every quarter.
Once, before any of this: add your site to Google Search Console. Free, five minutes, and it is how you raise your hand and say “come read me.” Then five to seven pages a week — start below that until the quality holds. AI cites content roughly 26% fresher than what ranks in ordinary Google.
One inconsistent phone number is worse than a missing profile — it tells the engine nobody here is sure who you are.
Share of relevant AI answers you appear in. The study measured Trustpilot; that Google and Yelp behave the same way is a fair bet, not a finding.
The whole transcript is fair game, so say the phrase out loud and put it in the title. “Best neighbourhoods in [your city]” is enough to start.
Real customer language.
The rules and four templates.
One answer-first page.
Facts, proof, legal risk.
Publish only after the manual win.
Context and skill from steps 03 to 06. Question bank from steps 07 to 09. It drafts; you still read every one. Publishing is one line of config away — and it waits until you have read a week of drafts.
Your whole OS is one folder on one laptop. GitHub is undo and a second copy. Vercel is how a thing you built ends up on the internet.
A real database. You want one when several people share the data, or when there is more of it than a person could sit and read.
Point it at hundreds of your documents instead of six context files. Most people are better served by six good files.
More autonomous harnesses, usually on a machine that never sleeps. More rope — and more that can go wrong while nobody is watching.
On your hardware, nothing leaving the building. Slower, free. Fine-tuning is almost never the answer — context and skills are.
A voice agent that picks up, qualifies, and books straight into your calendar. Probably the fastest money on this list.
Not one of these is required for what you built today. Each one solves a problem you will feel first — go looking when you feel it, not before.
Both prompts · this presentation · every tool we named · how to get started, step by step · and your link to try Relvara. Nothing to write down and nothing to sign up for right now — it is all coming to you.