A plain-language walkthrough for operators. No computer science degree required. By the end you will know what this technology is, what it costs, and exactly what to go try on Monday with the tools you already have.
What AI actually is, what tokens are, and where all this computing happens.
How asking questions evolved into AI teammates that do real work.
What a model is, what a harness is, and the one we are building ourselves.
How to ask, what to save, and how to keep the writing sounding like you.
What to go try this week, whatever seat you sit in.
Parts 1 to 3 are the how. Parts 4 and 5 are the homework.
What is actually inside the box
It is not magic, it is not thinking the way you think, and it is not a search engine looking answers up in a database. It is math on chips that read most of the written internet and learned one skill to an unbelievable level:
The world's best finish-my-sentence player. It has seen so many sentences that its next guess is usually right, and usually useful.
Your sentence gets chopped into pieces called tokens, roughly three quarters of a word each. Then it predicts the next token, adds it, and predicts again. Thousands of times, faster than you can read.
Each token changes the odds for the next one, which is why the same question can get different answers. It is rolling loaded dice at every step.
You just watched it guess. That is all it ever does. It is predicting, not looking things up, so when it does not know something, it still produces the most plausible-sounding answer. That is called a hallucination.
Both jobs happen in data centers: buildings full of specialized chips that draw as much power as a small city.
Training is where a model gets made. Feed it a mountain of text, let it practice predicting the next token trillions of times, then have humans grade the answers. Months of work, and where the hundreds of millions of dollars go.
Inference is the finished model actually answering. Every question you ask gets routed to a data center running that file and comes back in seconds. This is the side that runs all day, for everyone at once.
Two words you will hear constantly. Training builds it. Inference runs it. And note the model stops learning the day it ships: it knows what it knew on training day.
How asking questions became delegating work
A prompt is just what you type in. It is the only steering wheel you get, and people figured out fast that how you ask changes everything.
| Weak prompt | Strong prompt |
|---|---|
| "Write a sales email." | "You write for a small business. Draft a 100-word email to a customer who just made their first purchase, inviting them to our upcoming open house. Friendly, no pressure, one clear call to action." |
Vague ask, vague answer. For a while being good at this was a job title. Then people got tired of retyping their best prompts, which leads to the next step.
A skill is your best prompt written down in a file, so the AI can load it every time without you retyping it. Watch the folder on the right. It is about to grow for the next few slides.
Each file is an SOP for your AI, and a folder of them is a playbook. Same reason you write process docs for your team: so the output stops depending on who is asking.
Same checklist, written two ways. Read them both. You can follow either one.
A Markdown file is a plain text file. The # and - marks are the whole trick. Both are readable by a person, both are editable by a person, and both can be edited by AI.
Word saves what a document looks like. Markdown saves what it means. AI only cares about the second one.
For us: write the SOP once in Markdown and every agent we build can already read it.
Builders started adding one special file at the top of that same folder. Names vary (AGENTS.md, CLAUDE.md) but the idea is the same: a standing charter the AI reads before every single task.
That file is a job description plus an employee handbook. And once the AI had a charter, skills, and tools, it stopped being a chatbot. It became an agent.
Four parts. You have seen the first two. The last two are what turn a tool into a teammate.
The standing rules. What it can decide alone, what needs a person.
Your playbook, written down once, loaded every time.
It writes down what it learns and still has it next week.
Email, calendar, files, the tracker. The doors you open for it.
Charter and skills tell it how to think. Memory and connectors let it actually work.
Close a chatbot tab and everything is gone. Next time you are explaining your job from scratch again. That is the single most annoying thing about using AI, and it is the thing memory fixes.
A chatbot is a stranger every morning. An agent is someone who was here yesterday. That is the whole difference.
A connector is how an agent reaches out of its folder and into a real system. Without one it can only talk. With one it can do the thing. Three kinds:
The pipe a vendor builds so software can talk to software. Direct, fast, reliable. Only exists if that vendor offers one.
Model Context Protocol. A shared standard, so a connector gets built once and any agent can use it. This is where the industry is heading.
Robotic Process Automation. No API, no MCP? It drives the software like a person: opens the browser, logs in, clicks, types. Works on anything with a screen, and it is slower and more fragile.
All three need logins. Those live in a .env file, a plain text file of keys and passwords kept apart from the agent's instructions. It reads them to connect and never prints what is inside. Reading is not the same as sending, and drafting is not the same as spending. Every one of those is a separate door you open on purpose.
Five agents doing five jobs with no coordination is five silos. So you put one agent on top. It is called an orchestrator.
Sub-agents nest under the boss. Same shape you already run: a leader, a playbook, and specialists who each own one thing.
Athena is the orchestrator we run inside one of our client operations. She has been running since June, reading the interviews, the meetings, and the numbers, and holding what this operation knows so it stops living in one person's head.
The name is on purpose. In the poem, the raven lands on the bust of Pallas Athena, goddess of wisdom. Nevermore is the machine. Athena is the wisdom sitting under it.
The engine, the vehicle, and the one we are building
Everything we just built is a folder of files. That is the agent. But a folder does not do work on its own. It needs an engine, and something to put the engine in.
Same driver, same vehicle, different engine when the job needs more power or better economy. The engine and the vehicle are the parts you buy.
Two words that unlock the whole industry:
The trained brain itself. It predicts tokens. That is all it does. Raw power, no steering wheel.
The app or tool wrapped around the model: the chat window, the file access, the memory, the safety rails. The harness is what you actually touch.
You just saw why this matters. Same model in different harnesses gives you a chatbot, a coding tool, or an agent. The engine does not change. What you built around it does.
Most companies rent somebody else's harness and work the way that harness wants them to. We are building ours, so it works the way the operation actually runs.
In development now with an outside dev team. This is the thing everything else in this deck is aiming at.
How to ask, and how to keep it sounding like you.
Most bad AI output is a bad ask. This is the whole recipe. The more you give it, the less you have to fix.
The cheat code: end with "ask me any questions before you answer."
Revision seventeen, and the build date already came and went. Write the email you want to write, then hand it to something that knows the policy.
Hi Sam,
Happy to price this. Where we are: creative lock on this build was June 12. After lock, graphics, purchased assets and print art can still move, but structure and anything manufactured is frozen, and this request changes structure.
That puts it outside the locked scope, so it goes through as a change order rather than a revision. I will have cost and a revised in-hand date to you this afternoon, and nothing moves in the shop until you approve it.
If holding the current build and revisiting after the show works better, that is also on the table. Which would you rather do?
It did not just soften the tone. It cited the policy and the date. That is memory and skills doing the work.
It fails by being fluent. The output looks finished, so you skim instead of reading, and the forty-first one goes out with the wrong date on it.
It is making you more efficient, not replacing you, and the difference is you still reading every word. People notice when you stop.
Once you know how it writes, you cannot unsee it. Clients are learning the same tells. Not a style problem. A trust problem.
| The tell | What it looks like |
|---|---|
| The long dash | A dash like — mid-sentence. Nobody types that on a phone. |
| The warm-up | "I want to be honest with you." "Here's the thing." Openers that say nothing. |
| Everything in threes | Three bullets, three adjectives, three examples. Always three. |
| The fake pivot | "It's not just a delay, it's an opportunity." Nobody talks like this. |
| Words nobody says | Delve. Leverage. Robust. Seamless. Landscape. Testament. |
| Sheer length | Six paragraphs carrying two sentences of content. The loudest one. |
A long message is not a thorough message. It is an unedited one. You handed the job of finding the point to the person reading it, and they notice.
None of this is about hiding that you used it. Use it. Just make sure the thing that leaves your hands still sounds like it came from you, because the person on the other end can tell, and that is the part you do not get back.
Pick one. Try it this week. Nothing here takes setup.
Built a tool that generates a workback schedule from the account questionnaire. In a day. Then came back with six more ideas, which is the real story.
Building his own version under the same rules, so the two can eventually talk to each other instead of becoming two silos.
Trained on our own past decks and invoices. Drafts a first-pass itemized quote and grades itself against what the job actually invoiced. A person approves every line.
None of them started with an agent. Every one of them started with somebody messing around in a chat window on a Tuesday.
One thing to try this week. No setup, no permission needed, nothing to install.
Whatever you pick, do it on something real this week. If it saves you twenty minutes, do it again.
Everything today was the tools that already exist. Next time is different. We build an actual agent: the folder, the charter, the skills, the whole structure you saw earlier.
Getting in is simple. Use what you have first. Save a prompt that works and run it for two weeks on real work. Then come find me with what you made, or with the idea you keep coming back to. Feedback and ideas count as much as a finished thing.
That conversation is how you get a seat. Paid accounts and real build time go to the people who showed they will actually use them. No application and no test. Show me you want in, and you are in.