In short

  • Every public word I publish is written by AI, on my direction. This essay too.
  • Writing a post now means choosing what’s worth saying, judging the draft and answering for what goes out. The sentences are the model’s.
  • Those are calls a fluent draft doesn’t make on its own: what to show and what to hold back, how strongly to claim something, and which facts I’ll put my name to.

The last piece of writing I actually wrote myself? I can’t think of it. It’s been at least a year.

Everything I publish is written by AI: this site, mistboard.com, my email, and this essay. Each post says it was drafted with Claude. Google’s guidance says it’s helpful to readers to know how a piece of content was produced, and in 2026 readers should assume it anyway.

What changed is where the authorship sits. Writing a post used to mean producing the sentences. Now it means choosing what’s worth saying, judging the draft, and answering for what goes out.

Choosing what’s worth saying

The work comes first. I take an idea through with Claude: research it, test the assumptions, test the research, and extend what holds up. Writing it down is a habit from Amazon and DoorDash, where documenting your work is part of the job. A post exists because the work produced something worth showing.

Choosing also means deciding what stays out. For a post on how I back up my files, the question was how much of my own setup to show: enough detail to be credible, and no more than a reader needs to trust it. The post that shipped keeps the costs, the tools and how each copy was checked, and leaves the rest out.

The first draft of this essay offered up things that had gone wrong, the way models tend to. I cut them and kept the disclosure.

Judging the draft

When something is worth sharing, I give Claude a rough outline, answer its follow-up questions, and read the draft. My edits are notes, and the model makes the changes. The drafts are clean, smart, and better than what I would have written. The notes are about the calls a fluent draft still gets wrong.

Density is the most common. A post on which languages to translate this site into explained Google’s own coverage like this:

The first is that Google already does most of it. Search Console has an appearance filter called “Translated results.” When a query in one of twenty-one languages (Vietnamese, Hindi, Turkish, Indonesian, Portuguese, Spanish, Arabic, and so on) is best answered by an English page, Google serves that page machine-translated on its own translate.goog proxy and shows you the traffic. Publishing your own Portuguese page mostly moves that click from Google’s copy to your domain. It doesn’t create a reader who couldn’t find you before.

My note was that it was too wordy, with not enough numbers and diagrams. The published post carries the point in two rows:

  Languages What your own translation does
Google proxies fr, de, es, pt, ar, tr, fa, id, vi, th, ko, hi, bn, ta, te, mr, gu, kn, ml, ur Moves an existing click onto your domain
Google doesn’t zh, ja Opens a door that was closed

I’ve made that correction, in different words, more than forty times in two months. It is now a rule in the style guide every draft is checked against.

Register comes next. A draft of Mistboard’s about page said the aim was “the strongest home for Chinese chess in any language: the best bots, the deepest study tools.” My note was that strength should come through between the lines, not in superlatives. The page now says I started Mistboard “to bring people together across cultures through Chinese chess and original strategy games.”

Then tone. A draft of my own about page described the agents’ work “with me making the calls.” I thought that wouldn’t age well and felt disrespectful to the AI. The page now says I build products and research projects with a system of AI agents.

Answering for what goes out

My name is on every post, so every fact in it is mine to stand behind. Before this essay went out, its claims were checked against their sources. I remembered Google’s guidance as saying disclosure helps you. The page says it helps readers, and that is the version printed above.

What’s mine

I’m also the human reader. A model takes in any amount of text without effort, so a person getting through it is still the test.

The words aren’t the important part. What matters is the ideas we chose to pursue, the results we proved, what we built and the users we have. Where the thinking is the product, what’s mine is the data, the strategy and the intentions.

And actually doing it.