How I Write With AI Without Letting It Write for Me

I have argued on this site that the person who lets AI write for them ends up unable to think for themselves, and that the coach must read while the student writes. I believe that. I also owe the reader an admission, because for a while this site was an example of the thing I was warning about.

How the first sixty posts got written

When I started publishing here in early 2025, I did the thing everyone did that year. I had long conversations with a chatbot about what I believed, asked it to turn them into essays, converted the output to markdown, and posted it. The ideas were mine. The sentences, the structure, and the supporting evidence were not, and I did not check them the way I would have checked a paragraph I had typed myself.

This week I went back through every post with a stricter eye, and the audit was humbling. By my count there were more than a hundred quotation blocks attributing a round number to a famous institution, of which most could not be traced to any real source and several were simply wrong. Forty-five posts had the title rendered twice because the draft repeated it in the body. Two dozen posts used the emoji-bullet style that marks machine-written text from across the room. One link still carried a tracking parameter that announced where it had been pasted from. Thirteen posts were bad enough in tone or evidence that I unpublished them rather than repair them.

None of this was the tool’s fault. The tool did what I asked, which was to produce something that looked like an essay. I was the one who did not read it closely enough to notice that looking like an essay and being one are different things.

The pipeline I built and never ran

Here is the part that is more embarrassing than the drafts. In April 2025 I built a proper content pipeline for this site: research, outline, draft, edit, SEO, hero image, publish, each step a documented skill, with a state registry and human checkpoints between stages. Nine commands. Several thousand lines of instructions. It was, I thought, an AI-native writing process with editorial standards built in.

I never ran it. Not once. The registry that was supposed to track every post’s state stayed empty for seventeen months. The posts kept getting written the old way, because the old way was faster and the pipeline was a document, not a habit.

So this week I ran it, for real, on two of the new essays. Here is what happened.

The orchestrating agent could not be invoked at all, because its definition file was missing the header the tooling needs to register it. The step that tracks state wanted to write a field into each post’s frontmatter that the site’s content schema rejects, which would have failed the build. The skill that formats research mandated exactly the citation style I had just spent a day removing. The tag taxonomy in the SEO instructions did not include the site’s most-used tag. The workflow assumed drafts lived in a staging folder that the site does not use.

And once those were worked around, the pipeline was genuinely useful. The edit pass found a dead link to a publisher’s page and replaced it with the live one. It checked a figure I had written from memory, that online course completion rates were “around three percent,” against the primary source, and corrected it to what the paper actually shows: a fall from about five percent to about three over five years. It added two studies I had not known about, each with a link to the original PDF. The SEO pass took a title from sixty-two characters to fifty-eight, rewrote a description that was a hundred characters too long, replaced tags that no other post used with ones that did, and added two internal links where the prose already pointed at another essay.

The lesson is not that the pipeline was bad. It is that a process you have not run against the real thing is not a process. It is a hope. Every tool I build now gets run on a real post before I trust a word of its documentation.

The rules I write by now

Out of all this came a short set of rules. They are not sophisticated. They are just enforced.

AI researches. I decide what is true. I use it constantly to find sources, summarize papers, and check whether a figure I remember is real. It is very good at this. It is also confidently wrong often enough that nothing it tells me goes into a post until I have opened the source myself.

No statistic without a link to the primary source. Not a news article about the study. The study. If I cannot find it in about a minute, the sentence goes. This rule alone deleted around thirty claims from the archive, and I do not miss any of them.

Delete rather than re-source. When a claim turns out to be unsupported, the temptation is to go find a different number that supports the same point. That is how the original problem happened. If the argument needs a number to stand, and the number was invented, the argument was weaker than I thought and should be rewritten or cut.

AI reads. I write. For the essays themselves, the tool’s job is the one I described in The Feedback Loop Is the Teacher: read the draft and tell me where it is unclear, repetitive, or unsupported. It does not produce the draft. Paul Graham’s prediction in Writes and Write-Nots is that the world will divide into people who can still write and people who have outsourced it, and that writing is thinking. I spent a year on the wrong side of that line without noticing. The sentences on this site are mine now, with all the flaws that implies.

Keep the machine’s tells out. No emoji bullets. No one-line bold restatements after every paragraph. No quotation blocks attributed to institutions. No title repeated in the body. These are small, but they are how a reader knows within a paragraph whether a person was present.

Say what changed. Every post I revise gets a visible “last updated” date, and a post I retire gets a redirect to the nearest one I still stand behind, not a quiet deletion.

Why this matters beyond a blog

I care about this for a reason larger than my own archive. A lot of what I write about is how the next generation should learn, and I have argued that a student should use AI as a tutor and a coach while doing the hard part themselves. I cannot make that argument credibly from a site that did the opposite.

The same discipline applies to any team building with these tools. AI makes it cheap to produce something that looks finished. That is a gift and a trap in equal measure. The gift is speed. The trap is that “looks finished” and “is correct” have come apart, and the only thing that puts them back together is a person who reads every line as if their name is on it.

Mine is not on this site, by choice. I read it that way anyway now.