Compound judgment, not better artifacts

Half the drafts I ship now start as something a model wrote.

The prose is cleaner. The structure holds. The sources are lined up. And if you ask me why this version is stronger than the last one, I start reconstructing the argument from the chat log instead of from my own head. That is the tell. The artifact improved. My judgment of the next case stayed flat.

I care about the second one more than the first.

Polished output is the cheap win

Most AI use is optimized for removing friction. Ask, get an answer, move on. Nate B. Jones calls the opposite move friction maxing: deliberately hunting disagreement across models and trusted people so the first polished answer does not get to win by default. In his framing, disagreement is a rep. What survives several rounds of argument is rarely what the model handed you first.

I am not claiming that method as mine. I watch his loop, save the pieces that stick, and then ask a narrower question for my own work: what should compound while generation gets cheaper?

For me the durable output is not a better draft of this post. It is a sharper mental model of quality, taste, and failure that I can bring to the next newsletter pick, the next bot proposal, the next design critique. High agency, low authority already drew the boundary on the outside: generate freely, commit deliberately. This piece is about what should improve on the inside while that boundary holds.

The practice I actually run

When something feels off, I try to say what feels wrong before I ask for another polish pass. That sounds small. It is not. Naming the discomfort forces a claim I can later keep or discard. If I skip that step, the next model just sandpapers the surface until my unease disappears without ever being examined.

I also look for disagreement on purpose. Another model. A trusted person. Reality, when the thing can be checked against a file, a deploy, a published post. Agreeable refinement is easy to mistake for progress. Disagreement is slower and more useful.

And I keep the authority to throw the answer away. That is the unglamorous part. Seeking disagreement becomes theater if every round still ends with "sure, ship it." The point is not to collect opinions. The point is to leave with a judgment I can defend.

Emil Kowalski's Train Your Judgement exercises make a similar demand in a different domain: put into words why one animation feels better than another. Taste cannot be outsourced to a model that will happily produce motion that works and still feels mediocre. The act of saying why is the training.

The diagnostic I stole and kept

Nate asks a version of this that I now use as a personal test: after you use AI, do you feel more capable or less? My sharper version for myself is narrower.

Can I explain why my mind changed without asking a model to reconstruct the reason?

If yes, something compounded. If no, I probably just validated a better artifact.

I do not have a clean single case with a neat before and after stamped on it. What I notice instead is quieter. Some decisions around my daily AI working setup cannot be settled by sparring with a model all day. They need to sit. Sleep does work that another prompt cannot, or at least that is how it feels from the inside. Connections rearrange overnight. I am not sure how scientific that is. I am sure that time has been a better editor than another agreeable rewrite.

That is also where Dan Shipper's After Automation argument lands for me. Cheap competence does not erase the human job. It relocates it. Someone still has to stay ahead of the frame. Compounding judgment is one way of staying ahead without pretending the model never helped.

Where I am still the stick-together person

I asked myself where I might already be an efficient validator of AI output instead of a sharper judge of the next case. The honest answer, in this writing workflow, is that I do not feel like I have crossed that line yet.

AI does the heavy lifting: perspectives, draft structure, connections across my reading and older posts. I still decide what sticks. I still own the seams. That distinction matters to me the same way it mattered when I wrote about Making my newsletter smart. The picks stay mine. The last click stays mine. The joining tissue stays mine.

I am not romanticizing this. It is easy to slide from "I am stitching" into "I am rubber-stamping." The diagnostic is how I catch the slide. If I cannot say why this draft is better without reopening the chat, I am not done.

When this is the wrong tool

Not every task deserves friction maxing. An email subject line can just get better. Seeking disagreement can turn into model shopping with no clearer independent view at the end. The diagnostic also fails closed if you never change your mind at all.

Compounding judgment is overhead. Useful overhead on the decisions that teach you something about the next one. Wasteful overhead on work that only needs to leave the building once.

What I want to leave with

Serious AI use should compound my judgment across decisions, not only polish the current artifact.

Resist the first polished answer. Say what feels wrong. Seek disagreement from models, people, and reality. Keep the authority to discard. Then check whether you can explain the mind-change without outsourcing the explanation.

If the draft got better and you did not, that is information. Treat it that way.