Last week, I was in Boise for the Craft + Commerce conference.
I always love going to Craft + Commerce, but this year was extra special because I got an invite to a private mastermind that happened the day before the conference started.
Part of that mastermind was a 90-minute Q&A with James Clear, author of Atomic Habits.
If somehow you’ve never heard of Atomic Habits, it’s a phenomenal book that has sold over 25 million copies. It’s been at the top of the New York Times Bestsellers list for 5 years, and has been translated into 60 languages.
One of the reasons that Atomic Habits has been so successful is that James really obsessed over every printed word. It is Exhibit A for the case against AI slop.
So, of course, I had to ask him how he thought about using AI in the writing process. And in this newsletter, I want to share my takeaways from his detailed response.
AI tends toward the average
The first thing James pointed out was the thing that frustrates him most: ask AI for something genuinely original, and it gives you the obvious.
He described asking for “10 incredibly powerful but not well-known ideas about productivity.” And what came back? Stuff he already knew. The greatest hits. The conventional wisdom dressed up as insight.
Here's his theory on why:
Because of how the LLMs are designed, they’ve taken all the information, and then they give you the middle of the bell curve. They give you the average.
What James actually wants is the opposite.
I want the stuff that was a blog post in 2012 that was amazing, but then got buried in the archives of the internet and nobody talks about it anymore. I love that idea. But it’s hard to pull that stuff out of it right now.
Think about what that means for those of us building a PKM system.
The whole reason we capture ideas, save quotes, and take book notes is to build a curated collection of the non-average notes and ideas that are useful in our own creative process. When you follow your curiosity and capture the things that resonate (instead of the things you think might be important), your second brain becomes, by design, the tail end of the bell curve.
But that’s the part AI can’t hand you. You have to go collect it yourself.
It’s great with guardrails
If the open-ended creative stuff is where AI gets mushy, the opposite is true where the rules are tight.
The places where AI blows me away are things that really tighten down the guardrails. Coding, math – anything that has very specific rule-setting it will follow.
I’ve seen the same thing with my own vibe coding experiments. The key is that there’s a clear feedback loop and the guardrails are built into the medium itself.
The lesson: the more constrained and rule-bound the task, the more impressive AI becomes. The more open-ended and taste-driven, the more AI struggles and reverts back to the middle-of-the-bell-curve average.
It’s great for research
James mentioned that he has found real creative value in using AI for research. Specifically, he uses it to surface raw material he might not have found on his own.
He told a story about working on a writing project where he wanted to open with “an interesting airplane crash story that most people hadn’t heard.” So he asked AI for five lesser-known airplane crash stories. Two he had already heard about, but three he hadn't. And one of them sent him down a Wikipedia rabbit hole that surfaced exactly the kind of material he was looking for, which he’d never have stumbled onto otherwise.
On the research side, that can be really powerful.
An important thing to note here: he was clear on the job he was hiring the AI to do. In this case, it acted as a research assistant. The job was pulling threads faster than he could possibly do alone, which allowed him to do the human work of deciding which thread was worth following.
He uses it heavily when editing
Another place James leans on AI is editing.
He shared how he’ll often take a paragraph and ask AI to rework it, sometimes asking for one version in one writer’s voice and another version in a completely different voice.
And he usually doesn’t like any of them. At least, not as a whole. But they do help him piece together something of higher quality.
You end up getting like 8 puzzle pieces, and then you have to take it and put it together, and it has to go through your own filter and become your own language again. But it’s better at the end.
He’s not asking AI to write the paragraph. He’s asking it to generate options he can react to. He’s getting a bunch of puzzle pieces, most of which he throws away, but a few of which spark something. The output never ships as-is. It passes through his human filter and comes back out in his own voice.
He’s bullish on the future of AI
It would be easy to read all of this as James being a skeptic. He’s not.
When he talked about the creative limitations, he kept circling back to one caveat: “This is the worst it’s ever going to be.”
He’s genuinely bullish on where this is headed. Everything he described – the average answers, the difficulty surfacing obscure ideas – is a snapshot of the tools as they exist right now, not a verdict on what they’ll become. The bell curve might get a lot more interesting. The guardrails might widen.
So the goal isn’t to decide once and for all whether AI is “good” at creative work. It’s to build a working relationship with it now, while staying open to the fact that the most effective way to work with it is guaranteed to change over time.
The part AI can’t do (and probably shouldn’t)
He also described his “dream scenario.” The fantasy every writer has.
James said that for his next book, he has 500 or 600 pages of notes sitting in a messy Google Doc (James, let me introduce you to Obsidian 😉). And the single most painful part of his process is taking those 600 pages down to a finished 250-page draft.
He admitted:
If AI could magically do that, it would be a huge unlock for me.
But then he caught himself.
It can do it. I don’t know that it can do it well. And it wouldn’t be mine.
This is the heart of it. Because what he described next is the best definition of writing (really, of thinking) that I’ve come across in a long time.
That work of taking 600 pages down to 250, that is the thinking. It’s the choosing. You said this in two different ways – which one says it best? You have this idea and that idea – which one needs to go first? It’s 10,000 choices like that that end up leading to the finished book.
AI could make those cuts. But the cuts are how the voice gets amplified. It’s how taste gets made. And they wouldn’t be his cuts. AI wouldn’t make the same choices. And so it wouldn’t be the same book.
He even named the obvious counterargument:
Maybe you say, who cares, I just want the work done. But at least in the current state, something is lost. It’s not quite as good.
What this means for your PKM system
The reason “the thinking is the choosing” hit me so hard is that it describes exactly what we do with our notes in our PKM systems. Capturing is easier than it’s ever been. The hard, valuable, human part has always been the synthesis. Deciding what matters. Deciding what connects to what. Deciding which version of an idea says it best.
You can’t outsource the understanding.
That’s not a bug in the process that you should try to automate away. That’s the whole point. Those 10,000 small choices are what turn a pile of inputs into knowledge that’s actually yours.
This is also why I keep coming back to the idea that AI is a collaborator, not an author. James isn’t outsourcing his book to a machine. He’s using it to surface obscure research, to generate puzzle pieces he can react to, to sharpen a crappy first draft into a better one. At every step, the work still passes through his human filter. It still becomes his language. The judgment never leaves his hands.
If you let AI do the choosing for you, you don’t just get a generic result. You get the middle of the bell curve, and you skip the exact part of the process that was making you a better thinker.
The Bottom Line: AI makes the writing better, but not faster
I believe the most important thing James said about AI came while he was describing that editing process.
The AI didn’t make him faster. If anything, sorting through those options and turning them into his own language is more work than just writing the paragraph himself. What it did was raise his ceiling.
This quietly destroys the most common pitch for AI writing tools, which is speed. Save time. Write your newsletter in five minutes. Crank out a month of content before lunch.
Speed isn’t the point. Quality is.
Here’s how the most successful non-fiction author in the world thinks about using AI in his writing process:
It’s not less work… if I just took what it spit out it would be worse, but it can help me get to better work.
Not less work. Better work. And the better work only shows up because you’re still the one doing it.
