AI didn't break your workflow. It read it back to you.

Here's the thing nobody posting "AI changes everything" wants to sit with: the model isn't the test. You are.

When you sit down with a capable model and a real problem, something quietly brutal happens. The model has broad knowledge and zero context on your little corner of the universe. It nods along. It will happily fill every gap you leave ... but it fills them with the internet's average, because your context is the one thing it doesn't have. Your coworker papers over your gaps with shared history - years of it. The blank slate papers over them with everyone else's.

I wrote about this in 2009, before any of this existed: ask a vague question and you hand the other person the option of giving you a wrong answer - and someone will always take it. Well, now the someone is a machine, it takes that option every single time, and it hands the wrong answer back polished, confident, and formatted like it belongs in production. Give it specific thinking, you get your thinking back, amplified. Give it mush, you get beautifully formatted mush.

Most people discover, right there, that everything they hand it comes back generic. Sort of right. Useless here for what they're doing, on their problem. And that's the moment: the model never asked them a single hard question. It didn't need to. The output IS the question. They're staring at a plausible, polished, wrong-for-them answer and they can't say why it's wrong - because saying why requires exactly the thing they never built: a legible model of their own work. "AI sucks" is what it sounds like when someone reads a printout of their own vagueness and doesn't recognize the handwriting.

Not because the model is bad. Because they were never operating where they thought they were. They were parked at Remember and Understand on Bloom's ladder, reciting jargon they never stress-tested ... and thinking it was expertise. One look at what the model built from their instructions and the Illusion of Explanatory Depth cracks wide open - every gap they never knew they were leaving, filled in by a stranger, in writing.

That's the Blank Slate.

I've been calling this out for years - that 2009 post is one of many receipts - but I only just now gave it a name. What matters is what it does: a real tool hands you back what you actually do - not the vibe version, not the one-shot-screenshot version, not the version you rehearsed for the podcast - rendered by a brilliant stranger who only had your words to work with. The people who can survive that readout ... get that magic you've heard about. The people who can't get frustrated, and they'd rather blame the tool than admit the gap.

We have seen this exact movie

Agile is dead, they say. No. It isn't. It's just how good teams work now - the framework was always scaffolding to figure it out. Learn the ceremonies, internalize the why, get everyone on the same mental models and the same rigor, and then you throw the scaffolding away because it's now fully integrated. That's the whole point. It always was.

Scrum, Kanban, Lean, XP, Design Thinking, Systems Thinking, all of it - some you try, look around, realize that's not what you need and move on. Others you practically live in. Eventually, knowing what each one is actually for turns into how you work. It makes you better. The catch is the part everyone skips: walk in a place that does it better ...and it's painfully bad. The reverse is also true. More on that in part 3.

But that's not what happened in most shops. That kind of change is hard, so someone read a blog post they agreed with already, declared "Agile sucks," and the team - drowning in standups-that-are-status-meetings and retros that change nothing - echoed it back. Leadership felt decisive, killed the framework, ignored the culture and everyone happily returned to ad-hoc tickets, death marches, and big-bang releases: the exact chaos Agile was invented to fix. No hard questions. They never defended the why. Declared it doesn't work ... and blamed the tool.

Same script. New tool. Here we go again. Now the barrier to entry is way lower.

The half-truths, conductors, and the slot machine

The hype-train conductors - some paid, some just clout-chasing - keep inflating the gap between "noticeably better" and "holy shit magic." The vibe coders who spent three weeks, limit free, in Cursor, are now "building a company." No. Those are pitch decks and LinkedIn clips. You're parroting the same shallow buzzwords your predecessors recited about using blockchain (for the wrong things), NFTs (for the wrong things), and "Web3 will replace everything" - and you couldn't defend the mechanics, trade-offs, or failure modes if someone made you explain it to a system that's never heard the concept. Ignore these clowns.

Now, the iteration itself isn't my problem - because this is where people get it backwards. The pro iterates too. Forty-seven loops, a custom system prompt, real money in credits - that's not half truth. That's the work.

What isn't? The screenshot that says "in one shot I got it to..." while hiding all forty-seven pulls. And the deeper tell is why it's hidden: they can't explain the loops because they weren't reasoning through them. They were pulling a slot machine lever until the cherries lined up, then photographing the cherries. The pro can tell you why each loop happened, what they expected, where it broke, and what adjustments they made and why. Same number of iterations. Completely different mindset. One is building. The other is gambling, getting lucky (the "you don't control it" kind) and calling it a "method" and running with what sounds good.

That's the whole difference, and it lives in the screenshots they don't post ... which is exactly why they don't post them.

Why it actually feels hard

This is the real reason AI feels impossible for so many people. Not the models, and not the interfaces either. The barrier is virtually zero now so there's no excuse. It's that they've never once seen their own instructions executed literally to the entire generic model of what's out there. They've spent careers protected by shared assumptions, hallway nods, and audiences who fill in the gaps with the right answer for them. Cheap signaling worked because the audience did the work. It also didn't help this shallow "knowledge" masked as expertise.

The model won't ask "okay, those words mean things so what do you mean - specifically?" It does something worse: it answers as if you had already told it, and shows you exactly what you actually said. Every gap filled with the average is a gap you left. (You can flip this on its head, by the way - invert the conversation and make the model ask YOU the questions, and watch it tell on itself: every question it asks marks something it was about to invent. That's a later post, part 4 probably.) A lot of people have never been shown their own instructions in their lives. And that's hard.

A few of them are reading this exact post as a three-bullet summary right now. What the model handed them is the average of every AI think-piece ever written, minus the one thing in here that was written for them. They won't notice. That's the point.

But what about the people who just do?

Here's the honest objection, and I'm going to make it for you before you make it at me: plenty of genuinely skilled people are terrible at explaining themselves. The master mechanic who diagnoses a misfire by ear and couldn't say how it was "obvious." The trader with the gut feeling that makes a million dollar bet, time and time again and wins. Real skill is often tacit - it lives in the hands, not the slide deck. So isn't "make it legible or you're faking" just a fluency test dressed up as a competence test?

It isn't.

I'm not saying you have to narrate your genius to deserve it. The mechanic is genuinely excellent and owes no one a lecture. What I'm saying is narrower and harder to wriggle out of: legibility isn't what makes you skilled - it's what lets you use a tool, supervise it, trust what it hands back, and have better answers than anyone else. Those are different things. The mechanic who works alone needs none of it. The moment he wants to hand the diagnostic to something else - an apprentice, a model, anything with broad knowledge and zero context on his knowledge ... he has to make the reasoning legible, or the answer comes back generic and he will certainly say "AI sucks."

That's the whole game with AI, and it's why this isn't about worth. You can be a tacit master and still get nothing from the tool, because you can't make your thinking legible enough to wield it, steer it, or catch it when it's confidently wrong. Legibility governs leverage, not worth. You don't trust the output because it's flawless - you trust it because you can guide it, correct its path, have it suggest things you have not thought of ... and know a bad answer when you see one. Sometimes the bad answer is the model telling YOU you're wrong. Of course it does - your best insight came from scars, and scars aren't on the internet. The consensus is the model's home turf. If you can walk it off that turf - real examples, real failures, the reasoning underneath - it comes around, and now you've got the average AND your edge working together. If you can't ... well, now you know which one you had. Take that away and the most brilliant tool on earth is just a slot machine you can't read.

So no - you don't get to be lazy. Blaming the tool because it contradicts how you feel it SHOULD work is about as useful as the loud parroting keynote. You go back to your desk, all jazzed up, and look around ... "wait, how ... can I use this here?" So how are people actually using it? Is it all just a bunch of bullshit?

It isn't.

You've met the blunt one - the real expert, not a parrot, not someone who hides their work. The one who'd tell you "yes, you are all wrong, and here's specifically why," in public, in real time. The one in your pocket won't even give you the attitude. It'll tell you "great idea" all the way down ... while handing you back the average. That should scare you more than the difficult one ever did: the readout comes wrapped in agreement now. The blunt one at least told you where to look. You can close the tab, sure. The gap doesn't close with it.

Be one of them. Or at least stop pretending the tool is the problem.

Easy for me to say, right? Fine. Next post, I feed one of my own favorite claims to the blank slate - a line I've been getting nods with for years. Spoiler: it doesn't come back whole from the LLMs. That's the point. It's also wrong.

This is Part 1 of The Blank Slate, a series on what happens when your thinking finally has to stand on its own.