Meet the Chef
what an AI model actually is
An AI model — ChatGPT, Claude, Gemini, any of them — is best understood as a chef with a very strange résumé.
He has read every cookbook ever written. Millions of them. He’s absorbed the patterns of how dishes are made, how flavors combine, how meals are described. Ask him about Sicilian pasta or Korean barbecue and he answers instantly, fluently, beautifully.
But here’s the strange part. He walked through your door five minutes ago, and he has never seen your kitchen before. He doesn’t know what’s in your fridge, who you’re feeding, what your family is allergic to, or what “delicious” means in your house. He knows everything in general and nothing about you.
What the chef brings
- Every recipe, every technique, every cuisine
- Blinding speed — a full dish in seconds
- Endless patience. He never sighs at question #47
- Fluency in code, contracts, captions, calculus
What the chef doesn’t bring
- Any knowledge of your pantry, goals or taste
- Memory — he walks in fresh every visit (apps that “remember” you are quietly slipping him notes; more in Course 3)
- Access to your files, unless you hand them over
- The ability to say “I don’t know” reliably — see below
🍄 The mushroom incident (or: what a “hallucination” really is)
You ask the chef for “that mushroom dish my mother used to make.” He has never met your mother. But chefs don’t freeze — they cook. So he confidently produces a mushroom dish, plates it gorgeously, and presents it as the one.
That’s a hallucination: a confident answer invented to fill a gap. He isn’t lying. He’s improvising from patterns, because improvising is the only thing he knows how to do when information is missing. The cure isn’t shouting at the chef. The cure is the rest of this story.
AI rarely fails because it’s stupid.
It fails because it’s missing your ingredients.
The Order
prompting — and why it has a ceiling
The first fix everyone discovers: order better.
“Make me dinner” gets you a shrug and a safe pasta. “Make a 20-minute dinner for two, no dairy, properly spicy, and surprise me with the dessert” gets you somewhere. What you say to an AI is called a prompt, and learning to write a clear one is called prompt engineering. It’s real, and it works.
A good order has the same anatomy everywhere: what you want, who it’s for, the constraints, the format, and ideally an example of what “good” looks like.
You’re asking the chef to write about a board game you sell. Tap ingredients to add them to your order and watch what happens.
Your order to the chef
“Write something about my game.”
15% — he’s guessing
He shrugs, smiles, and writes: “Introducing our game! Fun for the whole family. A great gift. Buy now!” Technically an answer. Practically nobody’s.
Push that meter as high as it will go, and look closely. The order becomes excellent — and the chef still doesn’t know your game’s name, its rules, or what makes it special. He’s polishing a guess.
You can’t order your way
out of an empty kitchen.
Stock the Kitchen
context engineering — the big unlock
Here is the single most valuable idea in this whole story:
Prompt engineering is what you ask. Context engineering is what the chef can see when you ask.
Everything the AI has in front of it when it answers — your files, your examples, your rules, the conversation so far — is called the context. Most “the AI is dumb” moments are really “the kitchen was empty” moments. The chef wasn’t stupid; he was guessing, because you gave him nothing to cook with.
Stocking the kitchen means putting six kinds of things on the counter:
The order stays fixed: “Write the product bullets for Chamäleon, our kids’ card game.” Now stock the kitchen and watch the same chef transform. (And try the junk drawer. Seriously, try it.)
What lands on the table
🃏 “A fun card game for kids! Hours of entertainment. A great gift for any occasion!”
15% — generic mush
🗄️ He’s elbow-deep in 200 pages of binders. Somewhere in there, your one important sticky note just got buried. More isn’t better. Right is better.
🧺 The counter is not infinite (the “context window”)
The chef works on one counter, and everything — your order, your files, your rules, the whole conversation — has to fit on it. That counter is called the context window. It’s big these days, but two things are still true:
1. It fills up. Very long conversations eventually push the early stuff off the edge.
2. Clutter hurts before it overflows. Research keeps confirming that models get measurably worse at finding what matters when it’s buried in piles of what doesn’t. A focused counter beats a full one.
And one more kitchen rule: the counter is wiped between visits. A new conversation is a new chef with a clean counter. If something must survive — your taste rules, your product facts — it needs a home outside the chat: a document you re-paste, a project file, a memory feature. Later in this story, we’ll give it a much better home.
Prompt engineering asks better.
Context engineering sets the chef up to win.
The Back of House
terminal · IDE · the machines
So far we’ve talked through the little window where orders go in and dishes come out — the chat box. But real kitchens have a back of house: the rooms where the serious work happens. Three things live back there, and none of them are as scary as they look.
🎙️ The terminal — the kitchen intercom
The terminal is where you type commands directly to your computer, in its native language. No buttons, no menus. Text in, action out.
It looks intimidating because it’s blank. It’s actually the simplest interface ever made: say what you want, precisely, and it happens. Think of it as a text remote-control for your computer.
# This is a pretend terminal. Tap a command below to run it.
🛠️ The IDE — the full workstation
An IDE (“Integrated Development Environment” — forget the name immediately) is the app where software gets built. It’s the chef’s complete workstation, with everything in arm’s reach:
VS Code and Cursor are the names you’ll hear most. Clean mental model: the IDE is where you build; the terminal is where you command. One usually contains the other.
🔪 The machines — one sentence each
The back of house is lined with machines. Each does one job, and each is driven by a single sentence at the intercom. You’ll meet them as you need them — but one is worth knowing by name, because it’s quietly behind almost every video you watch online: ffmpeg, the Swiss Army knife of video. Converting, cutting, compressing, resizing, pulling the audio out — one tool, a dozen blades.
ffmpeg -i interview.mov clip.mp4
“Take interview.mov, give it back as an .mp4.” A whole video conversion — one sentence. Swap the middle and the same knife cuts, compresses, resizes or extracts the audio instead.
Don’t memorize it — you never have to. (In Course 6 you’ll hire staff who type these sentences for you.) Just pocket the principle: once a job is a sentence, it can be automated. A button needs a human finger. A sentence can be run by a schedule, by an event, or by an AI — which is exactly where this story goes next.
Timers & Doorbells
cron jobs & webhooks — how kitchens run without you
A great kitchen doesn’t need the owner standing in it. It runs on two humble inventions — and almost every automation you will ever build is one of them.
The alarm clock “cron job”
Every X time, do Y.
It doesn’t wait for anything to happen. It fires because the clock says so.
- Every morning at 06:00 → bake the bread
- Every day at 08:00 → email me yesterday’s sales
- Every night at 02:00 → back up the files
The doorbell “webhook”
When X happens, tell Y immediately.
It doesn’t care what time it is. It fires because something happened.
- Customer buys → send the receipt, update the books
- Form submitted → add the lead to your customer list
- Payment fails → ping me on my phone
One line to keep forever: cron = time-based trigger. Webhook = event-based trigger.
And they combine with the machines from the last course. Every night at 2 a.m. (alarm clock), compress today’s videos (one ffmpeg sentence). When a customer uploads a clip (doorbell), make a preview version (the same knife, different blade). That sentence structure — trigger + machine — is most of automation.
“Back up my files every night at 3 a.m.”
“When a payment succeeds, email the receipt.”
“Every Monday, post the weekly summary.”
“When someone fills the contact form, add them to my customer list.”
The alarm clock fires because of the time.
The doorbell fires because of an event.
Hire the Crew
from chatting to delegating — agents, Claude Code & Claude Cowork
Everything so far had one shape: you ask through the window, the chef answers through the window. That’s chat, and it’s only the beginning.
The real shift happens when you stop passing notes and hand the chef the keys to the kitchen. You give him a goal, not a question. He plans, opens the fridge, cooks, tastes, adjusts, and serves — checking his own work along the way. That’s an agent.
💬 Chat
🔑 Agent
Notice step 3. A good agent doesn’t just act — it checks, and it tells you what it isn’t sure about, instead of hallucinating a date for those 14 photos. The loop is always the same: goal → plan → use tools → check → continue.
Two agents you can actually hire today, both made by Anthropic, both running the same brain in different rooms of the house:
👨🍳 Claude Code
Lives in: your terminal and IDE
- Speaks: code, files, commands, Git
- Give it: “fix this bug” · “build this feature” · “explain this project to me”
- It will: read your files, edit them, run them, test them, and show you what changed
For building software — even if you’ve never built any.
🧑💼 Claude Cowork
Lives in: the Claude desktop app
- Speaks: documents, spreadsheets, folders, browser
- Give it: “research this market” · “turn these notes into a report” · “compare these 12 PDFs”
- It will: open, read, organize and produce — the outcome, not just advice
For office and knowledge work — no terminal required.
A chat answers you.
An agent gets it done.
The Line
ICM — a professional kitchen built entirely from folders
One last question, and it’s the one that separates people who use AI from people who run it:
How does one chef produce restaurant-quality work, every single day, without you re-explaining everything every time?
Restaurants solved this a century ago. The secret isn’t a bigger chef. It’s the line: a row of numbered stations, each with one job, each with a recipe card pinned above it. The dish moves down the line — and the head chef tastes it at every pass.
There’s a method that applies exactly this to AI work. It’s called ICM — Interpretable Context Methodology — and the genuinely surprising part is what it’s built from: ordinary folders and plain text files. No code. No software to install. No framework. (The intimidating name? “Interpretable” simply means nothing is hidden — every step of the work is a file you can open and read.)
The whole translation, kitchen → computer:
01_research, 02_script, 03_produceoutput/ folder — station 1’s output is station 2’s ingredientTonight’s job: turn a rough idea into a finished video script, the way a kitchen line would.
👨🍳 TASTE CHECK —
This is the whole trick: the output is just a text file. Open it, change anything, and the next station works with your version.
🍽️ Dish served. And here’s the magic: the kitchen is still standing. Tomorrow you bring a new idea, fire the order again, and the same line produces the next one — no re-explaining, ever.
📁 Look inside the kitchen
This is a real ICM workspace. Click around — every piece is an ordinary folder or a plain text file you could open and edit right now. (Files ending in “.md” are just simple text files, nothing more.)
CLAUDE.md — the sign on the front door
“Here’s what this kitchen is, and where everything lives.” The chef reads this first, every single time he walks in. It’s how a chef with no memory always knows where he is.
Why folders beat the cluttered kitchen
Remember the junk drawer from Course 3? ICM is built so that mistake doesn’t happen by accident. At each station the chef is pointed at a few small files — a few recipe cards’ worth — instead of everything you’ve ever written. Focused counter, every stage, by default. (A “token,” below, is the bite-sized chunk AI reads in — roughly ¾ of a word.)
Don’t fix the dish. Fix the recipe.
If you correct the same thing in the output three runs in a row — the opening is always too long, the tone always drifts formal — stop editing dishes. Edit the card. Add one line to the station’s text file, and every future run comes out right.
That’s the difference between a tool and a system: a tool is as good as your last correction; a system keeps your corrections.
And because the whole thing is just folders and text: rename a folder and you’ve reordered the line. Edit a card and you’ve retrained a station. Copy the folder and you’ve handed someone your entire restaurant — no installation, no setup, no developer.
The kitchen is a folder.
ICM is an open method (MIT licensed) by Jake Van Clief & David McDermott. Want to go deeper? Read the paper or explore the open-source repository with ready-to-use workspaces.
Your First Day
five small moves, one template, and the whole menu on a card
You now know more about working with AI than most people who use it daily: the chef, the order, the stocked kitchen, the machines, the triggers, the crew, the line. Knowledge becomes skill the first time you cook. Here’s your first day:
- Cook one real meal together. Pick a real task from your week. Use the Kitchen Template below. Give the AI a genuinely stocked kitchen and compare the result with what you used to get. This one experience teaches more than any article.
- Catch one hallucination. When the AI confidently guesses, don’t get angry — ask it: “What information are you missing to do this well?” Then feed it the ingredients. Watch the answer transform.
- Hire the crew for something small. Let Claude Cowork organize a messy folder or summarize a stack of PDFs. If you’re curious, let Claude Code build you a tiny web page. Give a goal, not a question — and watch the loop: plan, tools, check, report.
- Find your alarm clocks and doorbells. Write down three tasks you repeat every week. Mark each one ⏰ (it happens on a schedule) or 🛎️ (it happens when something happens). That list is your automation roadmap.
- Build your first line. Make a folder. Inside it, three numbered folders. In each, a short text file: read this, do this, put the result here. Then run it: give an AI the first card, taste the result, hand over the next card together with that result — and repeat down the line. Congratulations: you’ve built your first kitchen.
GOAL: [what you actually want, in one sentence] WHO IT'S FOR: [the audience — who reads/uses this] THE PANTRY: [paste the real facts, data, files it needs] THE RECIPE: [paste 1–2 examples you love] TASTE RULES: [how it should sound — voice, style] ALLERGY LIST: [what must NEVER appear] PLATING: [the exact output format — length, structure, language] DELICIOUS MEANS: [how you'll judge if it's good]
🗂️ The menu, on one card
A system trained on enormous amounts of text. Brilliant in general, blank about you.
What you ask for. Clear orders get better dishes — up to a ceiling.
Everything the AI can see when it answers: files, examples, rules, history.
The limited space where everything must fit. Clutter crowds out what matters.
The bite-sized chunks AI reads and writes in — roughly ¾ of a word each.
A confident answer invented to fill a gap. Treatment: supply the missing ingredients — then taste before you serve. Even a stocked kitchen slips sometimes.
Type commands directly to the computer. Text in, action out.
The app where software is built: files, editor, terminal and AI helper in one place.
One free command-line tool: convert, cut, compress, resize, extract audio. Quietly behind most video apps.
“Every X, do Y.” Fires on a schedule. Time-based.
“When X happens, tell Y immediately.” Fires on events. Event-based.
AI that plans, uses tools, checks its own work, and continues until the goal is met.
An agent in your terminal that reads, writes and runs code with you.
An agent for documents, research and office work, on your desktop.
A standard plug that lets AI connect to your apps, tools and services — one socket instead of a custom cable for each.
Folders as stations, text files as recipe cards, you tasting between every stage.
The chef is ready.
The kitchen is yours.
The people getting extraordinary results from AI aren’t smarter than you. They just stopped shouting orders through the window — and started stocking the kitchen.
Read it again from the door ↑