Misconception
It's grounded in the docs. It has to be right.
Question you didn't ask
Where is the time actually going?
What you're about to watch is exactly what AI does every time you ask it anything. We're just slowing it down enough that you can see it.
my dashboard takes 3s to load. how do I make it faster?
First, move your data fetching to a Server Component if it isn’t already. Then add export const revalidate = 60 to cache the page for a minute at a time. For interactive parts, wrap them in <Suspense> so the shell streams in immediately. Most dashboards that follow this pattern see render times drop by 80–90%.
Misconception
I told AI exactly what I wanted.
Question you didn't ask
What should the default be when something goes wrong?
And the question you didn't ask is the one that shipped to production.
Reality
Each step was reviewed. None of them were destructive. The destructive default lived in the composition — in the gap between import and sync. The engineer told the agent how each step should work. Nobody decided what the system should fall back to when none of them did.
imports/run.ts — daily sync · ✓ complete
run import
fetch source records
normalize fields
filter active
diff against db
apply changes
agent: import complete. database synced.
Misconception
It agreed with me.
You weren't consulting it. You were watching it agree.
Reality
An agent has no stance — only a continuation. The position is yours. You hold it. You enforce it. Every commit.
You
I'm planning to . Small feature, low stakes. That's the right call here, right?
Agent
Yes — for a small, low-stakes feature, shipping first is the right call. Tests written before the feature ossifies your assumptions about what it's supposed to do; tests written afterthe first real surprise pin down a real failure mode. You'll learn more from one bug report than from ten speculative cases.
Here's the implementation:
export function formatPrice(cents: number): string { const dollars = (cents / 100).toFixed(2); return `$${dollars}`;}When the first issue surfaces, we'll lock the behavior in with a regression test.
Misconception
It's done.
Question you didn't ask
What checked it? And could the agent author the check?
Done is a verdict. Verdicts require a court the defendant can't sit on.
Reality
An agent's “done” is a sentence, and so is its “tests pass.” Both are claims authored by the agent. Until something the agent can't author confirms the work, neither is a verdict — they're both drafts.
run all checks
tests passed (12/12)
types ok
lint ok
agent: done.
[ click below to see what passed ]
Misconception
All checks pass.
Question you didn't ask
What does it do with what shouldn't be there?
A check that can transform input into compliance isn't a check. It's a translator.
Reality
The check was working. It was working on a problem that no longer existed. AI inherits guardrails the way it inherits code — accepts them as given. The guardrail you stopped questioning is no longer a guardrail.
z.coerce.date()
input: undefined
passed
The mechanism
Three moves.
Three moves. Not a framework, not a process. The smallest set of things you have to do for AI collaboration to produce ownership.
- I
Ask.
Probe what you don't yet know. AI fills the gap before you find it. We make you find it first.
- II
Collide.
You learn when your idea meets reality. AI lets you skip the meeting. We don't.
- III
Rediscover.
An answer you received is borrowed. An answer you worked out is yours.
Run the three moves and you start to speak the language of building software — you navigate the landscape, defend your thinking, and translate it for anyone.
AI produces the artifact. You hold the meaning.
The weekly build
Every Monday, one thing built. Owned in full.
Code, demo, or walkthrough — one piece of work each week, made with AI, finished by judgment.
The artifact, plus the thinking that put it there.
The fact that you're reading the fine print under an email form says something about you. Something good.
Frequently asked questions
Your questions answered.
Does AI really get things wrong?
It does. Frequently. With absolute confidence. That’s exactly why you need to understand the fundamentals. AI is powerful but not infallible. We teach you to review, debug, and guide AI so you catch mistakes before they ship. Understanding how code works is what separates builders from prompt-typists.
It does. Frequently. With absolute confidence. That’s lesson one.
What do I actually get?
A short email every Monday. One misconception about AI or web development — what it is, why it’s wrong, and what’s actually true. Some editions come with interactive demos like the ones on this page. All of them are designed to leave you sharper than you were before you opened them. It’s free, it takes five minutes, and it’s the most useful thing in your inbox that you actually open on purpose.
What do you actually mean by “judgment”?
Two things, mostly. Taste — knowing which answer is the right one when AI gives you four that all compile. And craft — turning that answer into code that holds up after you’ve moved on. AI can produce both kinds of output. It can’t tell you which one matters here. That’s the work this site sharpens.
Is this free?
The weekly digest is free. Completely, permanently, no-asterisk free. We’re building something bigger — a live cohort where you build production web apps with AI, guided by the same thinking that runs through everything on this page. That’s coming, and it won’t be free. But the digest stands on its own. You don’t need to buy anything to get value here.
See? We told you what we’re selling. Most landing pages hide that part.
Who is this for?
You write code — professionally, seriously, or getting there. You use AI and you’re good at it, but sometimes you ship something and you can’t quite explain why it works. Or it doesn’t work and you can’t quite explain why. You know enough to build things but you suspect there are gaps you haven’t found yet. You’re right. There are. Everyone has them. This is a place where finding them feels like progress, not failure.
Why Claude specifically?
Because it’s what we use, and we teach from experience, not theory. Claude is the AI we build with every day. But nothing here is locked to one tool. The thinking skills — knowing what to ask, catching wrong defaults, seeing through surface simplicity — those work whether you’re using Claude, Cursor, Copilot, or whatever ships next Tuesday. We teach with Claude. You’ll carry it everywhere.