How I write

How I work with AI on this blog, and how I check what I publish. The argument is mine. The polish isn't entirely.

There is a paragraph in almost every essay on this site that I rewrote eleven times.

Not because the words were wrong. Because the thinking underneath was still soft. The first version was a thought. The eleventh was a position. Most readers will never know which is which.

That’s the point.

This page is about how that happens.

The loop

I sketch the spine of the essay first. Hook, turn, beats, closing image. Usually on paper. Sometimes in a markdown file with one line per paragraph. No AI yet.

Then I argue with Claude.

Section by section. I draft. It pushes back. I rewrite. It edits. I delete a paragraph it suggested and replace it with one I like better. It tells me the new one is weaker. Sometimes it’s right.

When we disagree about how a sentence sounds, I win. It has to sound like me. When we disagree about whether a sentence is true, nobody wins until I’ve checked.

The argument is mine. The shape is mine. The polish, the rhythm, the structural surgery: those were negotiated. The final call wasn’t. Every sentence that ships is one I chose to keep, and I answer for all of them.

What stays mine

The hook. Always. I write the opening myself. AI can attack it. It doesn’t get to write it.

The thesis. The reason the essay exists at all. The thing I’m willing to take heat for in the comments.

The personal beats. The confession sections. The specific things I did, screwed up, learned. AI cannot supply experience.

What AI does

Most of the time, AI isn’t writing for me.

It’s arguing with me.

It points at a paragraph and says: this doesn’t go far enough. Or: you said the same thing three times in different sentences. Or: you started with conviction and ended with hedging.

Sometimes it’s wrong.

Sometimes it’s painfully right.

AI can generate text. It cannot decide what is worth saying.

How I check what I publish

A model can make a wrong sentence sound finished. So the technical claims go through a separate pass.

Primary sources, not summaries. A claim about a paper traces to the paper. A claim about a tool traces to its repo, its docs, its changelog. If I can’t trace something, I cut it or label it as my own read.

If it can run, I run it. Commands, configs, API calls get executed before they get explained. When I can’t run something, say against a service I don’t have access to, the post says so.

The demos are built to be tested. The model in how LLM training works is a real one, 44 knobs, small enough to inspect. Backprop hands one of its Q knobs a gradient. Then the page measures that same gradient the slow way: nudge the knob up, rerun the whole network, nudge it down, rerun again, and take the slope. It shows you both numbers and the gap between them, live. At the starting settings, the gap is under a trillionth. That’s one gradient out of 44, not proof the whole model is right. But it’s a check you can watch, not a claim you have to take. The same post hands you a test that catches the model memorizing instead of learning. Press the button.

The draft gets attacked before you see it. I ask Claude and Gemini to tear the argument apart, not polish it. Two models can share the same blind spot, so this pass finds the weak claims. The sources and the running code decide them. On that same LLM post, the attack found five places where I’d claimed more than the evidence allowed. I had written that after the core training loop, the rest is scale. It isn’t. I had implied that big models stop memorizing. They don’t, and now the post cites the research that shows it. Each overclaim got narrowed, cited, or cut.

When I’m wrong, I say so. If you show me a mistake, I fix the post and add a dated note saying what changed. Quietly editing it away would be worse than the mistake.

Polish makes a claim sound true. Evidence decides whether it is.

Why this disclosure exists

Because you should know what you’re reading. Which parts are research, which parts are judgment, which parts are polish.

I’ve written essays on this site about AI replacing technical writers. I’ve written about taking yourself out of the equation. I can’t argue that AI is reshaping the craft while pretending it hasn’t reshaped mine.

The argument is mine. The polish isn’t entirely. You are reading the loop, not its absence.

If a specific essay used AI more heavily, less heavily, or not at all, a note in that essay will say so. This page is the default.

What I will not pretend

I will not pretend the sentences are wholly mine when they aren’t.

I will not pretend a model’s confidence counts as evidence. It doesn’t, for me or for it.

I will not pretend AI use is a moral failing that requires apology. The craft has always borrowed. Editors. Copy-editors. Beta readers. Writing groups. Claude is the latest member of that lineage. It happens to be available at 2 a.m. It is also the only member that can invent a source and sound certain about it. That’s what the checking is for.

The loop doesn’t guarantee better writing. It raises the bar I’m willing to ship at.

None of this is settled. Two years from now this page may read as quaint. I’ll rewrite it then.

It won’t be the first thing on this site I rewrote eleven times.